Blow-Up Theory For Elliptic PDEs in Riemannian Geometry

Blow-up Theory for Elliptic
PDEs in Riemannian
Geometry
This page intentionally left blank
Blow-up Theory for Elliptic

PDEs in Riemannian
Geometry
Olivier Druet
Emmanuel Hebey
eric
Robert
Fred
PRINCETON UNIVERSITY PRESS

PRINCETON AND OXFORD
c 2004 by Princeton University Press

Copyright
Published by Princeton University Press, 41 William Street, Princeton, New Jersey
08540
In the United Kingdom: Princeton University Press, 3 Market Place, Woodstock,
Oxfordshire OX20 1SY
All Rights Reserved
The publisher would like to acknowledge the authors of this volume for providing
the camera-ready copy from which this book was printed.
Library of Congress Cataloging-in-Publication Data
Druet, Olivier, 1976
Blow-up theory for elliptic PDEs in Riemannian geometry / Olivier Druet, Emmanuel Hebey, Frederic Robert.
p. cm
Includes bibliographical references.
ISBN: 0-691-11953-8 (pbk.: alk paper)
1. Calculus of Variations. 2. Differential Equations, Nonlinear. 3. Geometry,
Riemannian. I. Hebey, Emmanuel, 1964 II. Robert, Frederic, 1974 III. Title.
QA315.D78
515.353dc22
2004
2003064801
British Library Cataloging-in-Publication Data is available

Printed on acid-free paper.
pup.princeton.edu
Printed in the United States of America
10 9 8 7 6 5 4 3 2 1
Contents
Preface
Chapter 1. Background Material
1.1
1.2
Riemannian Geometry
Basics in Nonlinear Analysis
Chapter 2. The Model Equations

2.1
2.2
2.3
2.4
Palais-Smale Sequences
Strong Solutions of Minimal Energy
Strong Solutions of High Energies
The Case of the Sphere
Chapter 3. Blow-up Theory in Sobolev Spaces

3.1
3.2
3.3
The H12 -Decomposition for Palais-Smale Sequences

Subtracting a Bubble and Nonnegative Solutions
The De GiorgiNashMoser Iterative Scheme for
Strong Solutions
Chapter 4. Exhaustion and Weak Pointwise Estimates

4.1
4.2
Weak Pointwise Estimates

Exhaustion of Blow-up Points
Chapter 5. Asymptotics When the Energy Is of Minimal Type

5.1
5.2
Strong Convergence and Blow-up

Sharp Pointwise Estimates
Chapter 6. Asymptotics When the Energy Is Arbitrary

6.1
6.2
6.3
A Fundamental Estimate: 1
A Fundamental Estimate: 2
Asymptotic Behavior
vii
1
1
7
13
14
17
19
23
25
26
32
45
51
52
54
67
68
72
83
88
143
182
Appendix A. The Greens Function on Compact Manifolds
201
Appendix B. Coercivity Is a Necessary Condition
209
Bibliography
213
Preface
Let (M, g) be a smooth compact Riemannian manifold of dimension n 3. We
let (h ) be a sequence of smooth functions on M and consider equations like
g u + h u = u2
(E )
where g = divg is the Laplace-Beltrami operator, 2 = 2n/(n 2) is the

critical Sobolev exponent for the embedding of the Sobolev space H12 (M ) into
Lebesgues spaces, and u is required to be positive. Such equations have been the
target of investigation for decades. For instance, if h = ((n 2)Sg ) / (4(n 1)),
where Sg is the scalar curvature of g, then (E ) is the Yamabe equation.
We let (u ) be a bounded sequence in H12 (M ) of solutions of (E ) in the sense
that, for any ,
2 1
g u + h u = u
and u H12 where > 0 is independent of . We know from important work

developed in the 1980s how to describe the H12 -asymptotic behavior of the u s as
+. Let us assume, for instance, that the h s are uniformly bounded and
that they converge L2 to some function h . Up to a subsequence, the asymptotic
description with respect to the H12 -Sobolev space then gives that

u = u0 +
Bi + R
(I1 )
where u0 is a solution of the limit equation
g u + h u = u2
(E )
the sum in the right hand side of (I1 ) is a finite sum over i, Bi is a bubble,
obtained by rescaling fundamental positive solutions of the Euclidean equation

u = u2 1 , and the R s are lower order terms which converge strongly to 0
in H12 (M ). This asymptotic description provides a very nice H12 -theory for the
asymptotic behavior of solutions of equations like (E ). Let us assume now that
the h s converge C 0, to h for some 0 < < 1. An important issue in the study
of equations like (E ) is to get a theory in which the above asymptotic description
holds also in the C 0 -space, where pointwise estimates are involved. Developing
such a C 0 -theory is the purpose of these notes.
We wanted these notes to be as self-contained as possible. They should be accessible to graduate students and researchers in other fields. For the sake of clearness,
we decided to present our theory in the specific case of equations like (E ). However, the material we present in these notes is applicable to more general equations.
viii
PREFACE
In some aspects, this includes variants of equations (E ) involving the p-Laplacian.

It also includes variants of equations (E ) involving fourth order operators like the
square of the Laplacian.
Chapter 1 is devoted to background material in Riemannian geometry and nonlinear analysis on manifolds. The existence of Palais-Smale sequences for equations (E ), and of strong solutions of minimal or arbitrary energies to equations
(E ), is briefly discussed in Chapter 2. We present the H12 -decomposition theorem
in Chapter 3. Another construction, providing weak pointwise estimates, is presented in Chapter 4. Chapter 5 describes the C 0 -theory we mentioned above when
the solutions are of minimal type. The C 0 -theory in the general case of arbitrary
energies is presented in Chapter 6.
The authors thank Ellen Foos, Vickie Kearn, Alison Kalett, Jennifer Slater and
the Princeton University Press for their constant support, their efficiency, and the
wonderful job they did in the preparation of the manuscript.
The Authors
Paris, July 2003
Chapter One
Background Material
We recall in this chapter basic facts concerning Riemannian geometry and nonlinear
analysis on manifolds. For reasons of length, we are obliged to be succinct and partial. Possible references are Chavel [20], do Carmo [22], Gallot-Hulin-Lafontaine
[36], Hebey [43], Jost [50], Kobayashi-Nomizu [53], Sakai [65], and Spivak [72].
As a general remark, we mention that Einsteins summation convention is adopted:
an index occurring twice in a product is to be summed. This also holds for the rest
of this book.
1.1 RIEMANNIAN GEOMETRY
We start with a few notions in differential geometry. Let M be a Hausdorff topological space. We say that M is a topological manifold of dimension n if each point of
M possesses an open neighborhood that is homeomorphic to some open subset of
the Euclidean space Rn . A chart of M is then a couple (, ) where is an open
subset of M , and is a homeomorphism of onto some open subset of Rn . For
y , the coordinates of (y) in Rn are said to be the coordinates of y in(, ).
An atlas of M is a collection of charts (i , i ), i I, such that M = iI i .
Given an atlas (i , i )iI , the transition functions are
j 1
: i (i j ) j (i j )
i
with the obvious convention that we consider j 1
i if and only if i j = .
The atlas is then said to be of class C k if the transition functions are of class C k ,
and it is said to be C k -complete if it is not contained in a (strictly) larger atlas of
class C k . As one can easily check, every atlas of class C k is contained in a unique
C k -complete atlas. For our purpose, we will always assume in what follows that
k = + and that M is connected. One then gets the following definition of a
smooth manifold: A smooth manifold M of dimension n is a connected topological manifold M of dimension n together with a C -complete atlas. Classical
examples of smooth manifolds are the Euclidean space Rn itself, the torus T n , the
unit sphere S n of Rn+1 , and the real projective space Pn (R).
Given two smooth manifolds, M and N , and a smooth map f : M N from
M to N , we say that f is differentiable (or of class C k ) if for any charts (, ) and
)
the map
(,
of M and N such that f () ,
1
f : () (
)
k
is differentiable (or of class C ). In particular, this allows us to define the notion of
diffeomorphism and the notion of diffeomorphic manifolds.
CHAPTER 1
We refer to the above definition of a manifold as the abstract definition of a

smooth manifold. As a surface gives the idea of a two-dimensional manifold, a
more concrete approach would have been to define manifolds as submanifolds of
Euclidean space. According to a well-known result of Whitney, any paracompact
(abstract) manifold of dimension n can be seen as a submanifold of some Euclidean
space.
Let us now say some words about the tangent space of a manifold. Given M a
smooth manifold and x M , let Fx be the vector space of functions f : M R
which are differentiable at x. For f Fx , we say that f is flat at x if for some
chart (, ) of M at x, D f 1 (x) = 0. Let Nx be the vector space of such
functions. A linear form X on Fx is then said to be a tangent vector of M at x if
Given
Nx KerX. We let Tx (M ) be the vector space of such tangent
vectors.
T
(, ) some chart at x, of associated coordinates xi , we define x
x (M )
i x
by, for any f Fx ,

(f ) = Di f 1 (x) .
xi x

As a simple remark, one gets that the x
s form a basis of Tx (M ). Now, one
i x
defines the tangent bundle of M as the disjoint union of the Tx (M )s, x M . If
M is n-dimensional, one can show that T (M ) possesses a natural structure of a
2n-dimensional smooth manifold. Given a chart (, ) of M ,

Tx (M ),
x
is a chart of T (M ), where for X Tx (M ), x ,

(X) = 1 (x), . . . , n (x), X(1 ), . . . , X(n )
[the coordinates of x in (, ) and the components of X in (, ), that is, the
coordinates of X in the basis of Tx (M ) associated with (, ) by the process
described above]. By definition, a vector field on M is a map X : M T (M )
such that for any x M , X(x) Tx (M ). Since M and T (M ) are smooth
manifolds, the notion of a vector field of class C k makes sense. A manifold M
of dimension n is said to be parallelizable if there exist n smooth vector fields Xi ,
i = 1, . . . , n, such that for any x M , the Xi (x)s, i = 1, . . . , n, define a basis of
Tx (M ).
Given two smooth manifolds, M and N , a point x in M , and a differentiable
map f : M N at x, the tangent linear map of f at x (or the differential map of
f at x), denoted by f (x), is the linear map from Tx (M ) to Tf (x) (N ) defined, for
X Tx (M ) and g : N R differentiable at f (x), by

f (x) (X) (g) = X(g f ) .
By extension, if f is differentiable on M , one gets the tangent linear map of f ,
denoted by f . That is the map f : T (M ) T (N ) defined, for X Tx (M ),
by f (X) = f (x).(X). As one can easily check, f is C k1 if f is C k . Similar
BACKGROUND MATERIAL
to the construction of the tangent bundle, one can define the cotangent bundle of
a smooth manifold M as the disjoint union of the Tx (M ) s, x M. In a more

general way, one can define Tpq (M ) as the disjoint union of the Tpq Tx (M ) s,
where Tpq (Tx (M )) is the space of (p, q)-tensors on Tx (M ). Then Tpq (M ) possesses

a natural structure of a smooth manifold of dimension n 1 + np+q1 . A map
T : M Tpq (M ) is then said to be a (p, q)-tensor field on M if for any x

M , T (x) Tpq Tx (M ) . It is said to be of class C k if it is of class C k from
the manifold M to the manifold Tpq (M ). Given two manifolds M and N , a map
f : M N of class C k+1 , and a (p, 0)-tensor field T of class C k on N , one can
define the pullback f T of T by f , that is, the (p, 0)-tensor field of class C k on M
defined for x M and X1 , . . . , Xp Tx (M ), by

f T )(x) X1 , . . . , Xp = T f (x) f (x) X1 , . . . , f (x) Xp .
We now define the notion of a linear connection. Denote by (M ) the space of
differentiable vector fields on M . A linear connection D on M is a map D :
T (M ) (M ) T (M ) which satisfies a certain number of propositions. In local
coordinates, given a chart (, ), this is equivalent to having n3 smooth functions
kij : R, that we refer to as the Christoffel symbols of the connection in
(, ). They characterize the connection in the sense that for X Tx (M ), x ,
and Y (M ),

Y j
j
i
i
+ i (x)Y (x)
DX (Y ) = X (i Y )(x) = X
xi x
xj x
where the X i s and Y i s denote the components of X and Y in the chart (, ),
and for f : M R differentiable at x,

f
= Di f 1 (x) .
xi x
As one can easily check, the kij s are not the components of a (2, 1)-tensor field.
An important remark is that linear connections have natural extensions to differentiable tensor fields. Given a differentiable (p, q)-tensor field, T , a point x in M ,
X Tx (M ), and a chart (, ) of M at x, DX (T ) is the (p, q)-tensor on Tx (M )
defined by DX (T ) = X i (i T )(x), where
j1 ...jq
p

Ti1 ...ip
j ...j

j1 ...jq
i T (x)i11...ipq =
iik (x)T (x)i1 ...ik1 ik+1 ...ip
xi
x
k=1
q
j ...j
jik (x)T (x)i11...ipk1
jk+1 ...jq
k=1
The covariant derivative commutes with the contraction in the sense that

DX Ckk12 T = Ckk12 DX (T )
where Ckk12 T stands for the contraction of T of order (k1 , k2 ). Given a (p, q)-tensor
field of class C k+1 , T , we let T be the (p + 1, q)-tensor field of class C k whose
components in a chart are given by

j1 ...jq

j1 ...jq
T i ...i
= i1 T i2 ...ip+1 .
1
p+1
CHAPTER 1
By extension, one can then define 2 T , 3 T , and so on. For f : M R a smooth

function, one has that f = df and, in any chart (, ) of M ,
2

2
f
f
kij (x)
f (x)ij =
xi xj x
xk x
where
2f
xi xj

x

2
= Dij
f 1 (x) .
In the Riemannian context, f is called the Hessian of f and is sometimes denoted by Hess(f ).
The torsion T of a linear connection D can be seen as the smooth (2, 1)-tensor
field on M whose components in any chart are given by the relation Tijk = kij kji .
One says that the connection is torsion-free if T 0. The curvature R of D can
be seen as the smooth (3, 1)-tensor field on M whose components in any chart are
given by the relation
l
=
Rijk
lji
lki
l
+ lj
ki k ji .
xj
xk
l
l
As one can easily check, Rijk
= Rikj
. Moreover, when the connection is torsionfree, one has that
l
l
l
Rijk
+ Rkij
+ Rjki
= 0 and
(i R)lmjk + (k R)lmij + (j R)lmki = 0 .

Such relations are referred to as the first Bianchi and second Bianchi identities.
We now discuss Riemannian geometry. Let M be a smooth manifold. A Riemannian metric g on M is a smooth (2, 0)-tensor field on M such that for any
x M , g(x) is a scalar product on Tx (M ). A smooth Riemannian manifold is a
pair (M, g) where M is a smooth manifold and g a Riemannian metric on M . According to Whitney, for any paracompact smooth n-manifold there exists a smooth
embedding f : M R2n+1 . One then gets that any smooth paracompact manifold
possesses a Riemannian metric. Just think of g = f , where is the Euclidean
metric. Two Riemannian manifolds (M1 , g1 ) and (M2 , g2 ) are said to be isometric
if there exists a diffeomorphism f : M1 M2 such that f g2 = g1 .
Given a smooth Riemannian manifold (M, g), and : [a, b] M a curve of
class C 1 , the length of is

b
d
d
L() =
dt
g((t))
,
dt t
dt t
a

d
where ( d
dt )t T(t) (M ) is such that ( dt )t f = f (t) for any f : M R
differentiable at (t). If is piecewise C 1 , the length of may be defined as the
sum of the lengths of its C 1 pieces. For x and y in M , let Cxy be the space of
piecewise C 1 curves : [a, b] M such that (a) = x and (b) = y. Then
dg (x, y) = inf L()
Cxy
BACKGROUND MATERIAL
defines a distance on M whose topology coincides with the original one of M . In

particular, by Stones theorem, a smooth Riemannian manifold is paracompact. By
definition, dg is the distance associated with g.
Let (M, g) be a smooth Riemannian manifold. There exists a unique torsion-free
connection on M having the property that g = 0. Such a connection is the LeviCivita connection of g. In any chart (, ) of M , of associated coordinates xi , and
for any x , its Christoffel symbols are given by the relations

gmj
1
gmi
gij
k
g(x)mk
ij (x) =
+
2
xi x
xj x
xm x
where the g ij s are such that gim g mj = ij . Let R be the curvature of the LeviCivita connection as introduced above. One defines
1. the Riemann curvature Rmg of g as the smooth (4, 0)-tensor field on M
,
whose components in a chart are Rijkl = gi Rjkl
2. the Ricci curvature Rcg of g as the smooth (2, 0)-tensor field on M whose
components in a chart are Rij = Rij g , and
3. the scalar curvature Sg of g as the smooth real-valued function on M whose
expression in a chart is Sg = Rij g ij .
As one can check, in any chart,
Rijkl = Rjikl = Rijlk = Rklij ,
and the two Bianchi identities are
Rijkl + Riljk + Riklj = 0 ,

i Rmg jklm + m Rmg jkil + l Rmg jkmi = 0 .
In particular, the Ricci curvature is symmetric, so that in any chart Rij = Rji .
Given a smooth Riemannian manifold (M, g), and its Levi-Civita connection D,
a smooth curve : [a, b] M is said to be a geodesic, if for all t,

d
D d
= 0.
dt
dt t
This means again that in any chart, and for all k,

k
(t) + kij (t) i (t) j (t) = 0 .
For any x M , and any X Tx (M ), there exists a unique geodesic : [0, ]

M such that (0) = x and ( d
dt )0 = X. Let x,X be this geodesic. For > 0 real,
x,X (t) = x,X (t). Hence, for X small, where stands for the norm in
Tx (M ) associated with g(x), one has that x,X is defined on [0, 1]. The exponential
map at x is the map from a neighborhood of 0 in Tx (M ), with values in M , defined
of
by expx (X) = x,X (1). If M is n-dimensional and up to the assimilation

Tx (M ) to Rn via the choice of an orthonormal basis, one gets a chart , exp1
x
of M at x. This chart is normal at x in the sense that the components gij of g in this
CHAPTER 1
chart are such that gij (x) = ij , with the additional property that the Christoffel
symbols kij of the Levi-Civita connection in this chart are such that kij (x) =
0. The coordinates associated with this chart are referred to as geodesic normal
coordinates.
Given a smooth Riemannian n-manifold (M, g), one can define a natural positive
Radon measure on M . In particular,
can be ap

the theory of the Lebesgue integral
plied. For some atlas of M , i , i iI , we shall say that a family j , j , j jJ

is a partition of unity subordinate to i , i iI if the following holds:
(i) (
j )j is a smooth partition of unity subordinate to the covering (i )i ,
(ii) j , j j is an atlas of M , and
(iii) for any j, suppj j .

As one can easily check, for any atlas i , i iI of M , there exists a partition

of unity j , j , j jJ subordinate to i , i iI . One can then define the Riemannian measure as follows:
Given a continuous map f : M R with compact

support, and an atlas i , i iI of M ,

j |g|f 1
f dv(g) =
j dx
M
jJ
j (j )

where j , j , j jJ is a partition of unity subordinate to i , i iI , |g| stands
of g in
for the determinant of the matrix whose elements are the components
j , j , and dx stands for the Lebesgue volume element of Rn .One can
prove
that such a construction does not depend on the choice of the atlas i , i iI and

the partition of unity j , j , j jJ .
The Laplacian acting on functions of a smooth Riemannian manifold (M, g) is
the operator g whose expression in a local chart of associated coordinates xi is
2

u
u
.
kij
g u = g ij
xi xj
xk
For u and v of class C 2 on M , one then has the following formula for integration
by parts:

(uv) dv(g) =
u g v dv(g)
g u vdv(g) =
M
where (, ) is the scalar product associated with g for 1-forms.

Coming back to geodesics, one can define the injectivity radius of (M, g) at
some point x, denoted by ig (x), as the largest positive real number r for which
any geodesic starting from x and of length less than r is minimizing. One can then
define the (global) injectivity radius by
ig = inf ig (x) .
xM
One has that ig > 0 for a compact manifold, but it may be zero for a complete
noncompact manifold. In a similar way, one can define the cut locus Cx of x,

where
Cx is a subset of M , and prove that Cx has measure zero, that ig (x) = dg x, Cx ,
BACKGROUND MATERIAL
and that expx is a diffeomorphism from some star-shaped domain of Tx (M ) at 0

onto M \Cx . In particular, one gets that the distance function r to a given point is
differentiable almost everywhere, with the additional property that |r| = 1 almost
everywhere.
As is well known, curvature assumptions may give topological and diffeomorphic information on the manifold. A striking example of the relationship that exists between curvature and topology is given by the Gauss-Bonnet theorem, whose
present form is actually due to the works of Allendoerfer [2], Allendoerfer-Weil
[3], Chern [21], and Fenchel [35]. One has here that the Euler-Poincare characteristic (M ) of a compact manifold can be expressed as the integral of a universal
polynomial in the curvature. For instance, when the dimension of M is 2,
1
(M ) =
Sg dv(g) ,
4 M
and when the dimension of M is 4, as shown by Avez [8],

1
1
1 2
2
2
(M ) =
|Wg | + Sg |Eg | dv(g) ,
16 2 M 2
12
where | | stands for the norm associated with g for tensors, and where Wg and Eg
are, respectively, the Weyl tensor of g and the traceless Ricci tensor of g. In a local
chart, the components of Wg are

1
Rik gjl + Rjl gik Ril gjk Rjk gil
n2

Sg
gik gjl gil gjk
+
(n 1)(n 2)
Wijkl = Rijkl
where n stands for the dimension of the manifold. As another striking example of
the relationship that exists between curvature and topology, one can refer to Hamiltons theorem [39]: any three-dimensional, compact, simply connected Riemannian
manifold of positive Ricci curvature must be diffeomorphic to the unit sphere S 3 .
Conversely, by recent results of Lohkamp [59], negative sign assumptions on the
Ricci curvature have no effect on the topology, since any compact manifold possesses a Riemannian metric of negative Ricci curvature. As another example of the
relationship that exists between curvature and topology, we refer to the well-known
sphere theorem of Berger [10], Klingenberg [51, 52], Rauch [61], and Tsukamoto
[75].
1.2 BASICS IN NONLINEAR ANALYSIS
Given a smooth compact n-dimensional Riemannian manifold (M, g), one easily
defines the Sobolev spaces Hkp (M ), following what is done in the more traditional
Euclidean context. For instance, when k = 1 and p > 1, one may define the
Sobolev space H1p (M ) as follows: for u C (M ), we let
uH1p = up + up
CHAPTER 1
where .p is the Lp -norm with respect to the Riemannian measure dvg . We then
define H1p (M ) as the completion of C (M ) with respect to .H1p . A similar
definition holds for Hkp (M ), with
uHkp =
k
i up .
i=0
H1p
are that Lipschitz functions on M do belong to the

Very useful properties of
p,and that if u H1p (M ) for some p, then we have
Sobolev spaces H1p (M ) for all

p
that |u| H1 (M ) and |u| = u almost everywhere.
As for bounded open subsets of the Euclidean space, the Sobolev embedding
theorem (continuous embeddings), and the Rellich-Kondrakov theorem (compact
embeddings), do hold. In order to fix ideas, we let k = 1 and p = 2. Let
2 =
2n
n2
be the critical Sobolev exponent. Then, for any q [1, 2 ], H12 (M ) Lq (M ) and
this embedding is continuous, with the property that the embedding is also compact
if q < 2 . The Sobolev inequality corresponding to the continuous embedding

H12 (M ) L2 (M ) is as follows: For any u H12 (M ),
u2 Au2 + Bu2
where A and B are positive constants independent of u, but that may depend on
the manifold. Another very useful inequality is the so-called Poincare inequality.
When dealing with H12 , it reads as the existence of a positive constant A such that,
for any u H12 (M ),
u u2 Au2
where
u=
1
Vg
udvg
M
is the average of u, and Vg the volume of M with respect to g. In this particular

case, thanks to the Rayleigh characterization of the first nonzero eigenvalue of the
Laplacian, A may be taken to be the inverse of the square root of this eigenvalue.
Combining the Sobolev inequality and the Poincare inequality, one gets the socalled Sobolev-Poincare inequality: for any u H12 (M ),
u u2 Au2 ,
where A is a positive constant, independent of u as usual.
A very useful notion concerning Sobolev embeddings, which appeared to be
crucial in many problems like the Yamabe problem, is that of best constants. In the
particular case k = 1 and p = 2, the Sobolev inequality in the Euclidean space
reads as

12
21
2
2
A
|u| dx
|u| dx
.
Rn
Rn
BACKGROUND MATERIAL
The best constant A in this inequality, which we denote by Kn , is

4
Kn =
2/n
n(n 2)n
where n is the volume of the unit n-sphere. Taking A = Kn in the above inequality, we get what we refer to as the sharp Euclidean Sobolev inequality. Its extremal
functions are known. They are expressed as

1 n2
u,a, (x) = + |x a|2
where > 0, R, and a Rn are arbitrary. For compact manifolds, we
consider the Sobolev inequality
u22 Au22 + Bu22 .
As is easily checked, any such constant A must satisfy A Kn2 . Conversely, we
refer to Hebey-Vaugon [47, 48], it can be proved that there exists a positive constant
B such that for any u H12 (M ),
u22 Kn2 u22 + Bu22 .
More developments on Sobolev spaces, Sobolev inequalities, and the notion of best
constants are in Druet-Hebey [29] and Hebey [44].
Let (M, g) be a smooth compact Riemannian manifold. The equations we will
be interested in are basically of the form
g u + au = f
where a, f are given functions on M . A function u H12 (M ) is said to be a weak
solution of this equation if, for all H12 (M ),
u, g dvg +
a(x)udvg =
f (x)dvg .
M
Regularity results for this equation do hold. They are similar to the more traditional
ones expressed in the Euclidean context (regularity is a local notion, so this is not
very surprising). The regularity result we will mostly use is the following: if a is
smooth, and f Hkp (M ) for some k N and p > 1, then a weak solution u to
p
(M ). In particular, it follows from this result and the
the above equation is in Hk+2
Sobolev embedding theorem, that when f is smooth u is also smooth. Needless to
say, the bible for such topics is the exhaustive Gilbarg-Trudinger [37]. A simpler,
but very nice reference is the lecture notes [41] by Han and Lin.
In parallel with regularity, the very useful maximum principles hold for the
Laplacian on Riemannian manifolds. A currently used form is as follows: if a
nonnegative u C 2 (M ) is such that, for any x M ,
g u(x) u(x)f (x, u(x))
for some continuous functions f : M R R, then either u is everywhere
positive, or u is the zero function. This easily follows from the Hopf maximum
principle, as usually stated.
10
CHAPTER 1
In order to end this section, we give a basic example of a possible use of the
above results. We discuss here the existence (and uniqueness) of a solution u to the
Laplace equation
g u = f
on a compact Riemannian manifold (M, g). Although not necessary, we assume
here for convenience that f : M R is smooth. Integrating the Laplace equation,
one sees that a necessary condition for the existence of a solution is that
f dvg = 0 .
M
The elementary result we wish to briefly discuss

here is that the Laplace equation

possesses a smooth solution if and only if M f dvg = 0. Moreover, the solution is
unique, up to the addition of a constant. In order to prove this claim, we proceed as
follows. As already mentioned, the condition that f is of null average is a necessary
condition. We prove now that it is also a sufficient condition. Let

2
udvg = 0 and
f udvg = 1
H = u H1 (M ) s.t.
M
and
= inf
uH
|u|2 dvg .
Clearly, H = . Consider a minimizing sequence (ui ) H for so that ui H

for all i, and
|ui |2 dvg = .
lim
i+
By the Poincare inequality we discussed above, there exists A > 0 such that, if
u H12 (M ) is of null average, then
u2 dvg A
|u|2 dvg .
M
It easily follows that the ui s are bounded in H12 (M ). Since H12 (M ) is a reflexive
space, and the embedding H12 (M ) L2 (M ) is compact by the Rellich-Kondrakov
theorem (even when n = 2, noting that H12 H1q for q < 2), there exists a function
u H12 (M ) and a subsequence (ui ) of (ui ), such that
(1) (ui ) converges weakly to u in H12 (M ) and
(2) (ui ) converges to u in L2 (M ).
By (2), u H. By (1), and a basic property of the weak limit (the norm of a weak
limit is less than or equal to the infimum limit of the norms of the sequence), we
get that
|u|2 dvg .
M
Hence,
|u|2 dvg =
11
BACKGROUND MATERIAL
and is attained. By a well-known theorem of Lagrange, this gives the existence

of two constants and , the Lagrange multipliers, such that, for all H12 (M ),
u, g dvg =
dvg +
f dvg .
M
Taking = 1, one gets that = 0. Taking = u, one gets that = . Hence,

u is a weak solution of the Laplace equation. By standard regularity results, u is
smooth. The function 1 u is then the solution we were looking for. The proof
of uniqueness is then very simple. If u and v are two solutions of the Laplace
equation, then g (v u) = 0. Multiplying this relation by v u and integrating
over M gives that
|(v u)|2 dvg = 0 .

M
Hence, v u is constant, and this ends the proof of the above claim. As is easily
checked, everything in this proof comes from the compactness of the embedding
H12 L2 . The equations that we discuss below involve critical (noncompact)
Sobolev embeddings.
Chapter Two
The Model Equations
let (h ) be a sequence of smooth functions on M , and consider equations like
g u + h u = u2
(E )
where g = divg is the Laplace-Beltrami operator, 2 = 2n/(n 2) is the

critical Sobolev exponent for the embedding of the Sobolev space H12 (M ) into
Lebesgues spaces Lp (M ), and u is required to be positive. By standard regularity theory, as developed by Gilbarg-Trudinger [37], and thanks to the maximum
principle, if u H12 (M ) is a nonnegative solution of (E ), then u is smooth and
either u > 0 everywhere or u 0. As already mentioned, thanks to the Sobolev
embedding theorem, H12 (M ) Lp (M ) for p 2 , this embedding being compact
if p < 2 and not compact if p = 2 .
Equations like (E ) arise naturally in several problems. This is the case for the
Yamabe problem (see Schoen [66]) but also for sharp constant problems in Sobolev
inequalities. Possible survey articles on the Yamabe problem are Hebey [42], Lee
and Parker [54], and Schoen [69]. Possible monographs on sharp constants problems are Druet and Hebey [29], and Hebey [44]. We refer also to Druet [27], Hebey
[45], Hebey and Vaugon [49], and Schoen [67, 70] for extensions of these problems
where understanding equations like (E ) is necessary. The reference Hebey [45]
discusses the notion of the energy function. The reference Hebey-Vaugon [49] discusses the notion of critical functions. Among other topics, the references Druet
[27] and Schoen [67, 70] discuss compactness results.
We are concerned in these notes with the blow-up behavior of sequences of solutions of equations like (E ). A very nice H12 -theory for the blow-up behavior
of such sequences was developed in the 1980s. Among other possible references,
we refer to Brezis-Coron [12, 13], Lions [58], Sacks-Uhlenbeck [64], Schoen [67],
Struwe [73], and Wente [76]. A partial survey of the subject is given by EkelandGhoussoub [34]. The H12 -theory as in Struwe [73], dealing with Palais-Smale
sequences, is presented in Chapter 3 below. Atkinson-Peletier [5], with arguments from ordinary differential equation theory, and Brezis-Peletier [16], with
arguments from partial differential equation theory, have been concerned with the
description of the pointwise behavior of sequences of solutions of equations like
(E ), dealing with radially symmetrical solutions u of the semi-critical equations

u = u2 1 on the unit ball of the Euclidean space. We refer also to AdimurthiPacella-Yadava [1] and Robert [62, 63] for more recent developments. Roughly
speaking, a bubble is the rescaling of a fundamental positive solution of the critical
14
CHAPTER 2
Euclidean equation u = u2 1 . An estimate like the one we present in Chapter

5, stating that solutions of minimal energy of equations like (E ) are controlled
from above by a standard bubble, appeared in Han [40], dealing with solutions u

of the equations u = u2 1 on bounded open subsets of the Euclidean space,
in Hebey-Vaugon [47, 48], dealing with (E ) and arbitrary Riemannian manifolds,
and in Li [55, 56], dealing with equations like (E ) on the unit sphere. Other possible references are Chang-Yang [18], Chang-Gursky-Yang [19], Druet [24, 25],
Druet-Hebey [30], Druet-Robert [33], and Li-Zhu [57]. Weak estimates, like the
ones we present in Chapter 4, are reminiscent of the definition by Schoen [68] of
isolated blow-up points. We refer also to Schoen-Zhang [71]. Such estimates have
been developed by Druet [23, 26] when studying sharp constant problems.
We briefly discuss in this very basic introductory chapter the existence of PalaisSmale sequences for (E ), and the existence of strong solutions for (E ) of minimal or arbitrarily high energies. More sophisticated examples can be found in
Druet-Hebey [31].
2.1 PALAIS-SMALE SEQUENCES
Let h be a smooth function on M . Following standard terminology, we say that an
operator like Lg = g + h is coercive if there exists C > 0 such that
(Lg u)udvg
(2.1.1)
u2H 2 C
1
H12 (M ),
where the right hand side of this equation has to be understood

for all u
in the distributional sense, so that

(Lg u)udvg =
|u|2 + hu2 dvg .
M
Equation (2.1.1) just says that the H12 -norm of a function u is controlled by the
energy of u with respect to Lg . It is easily checked that, if h > 0 everywhere, then
Lg is coercive. As another remark, the existence of a u > 0 solution of (E )
implies that the operator L
g = g + h in the left hand side of equation (E ) is
coercive. From now on, we let Ig be the functional defined on the Sobolev space
H12 (M ) by

1
1
1
|u|2 dvg +
h u2 dvg
|u|2 dvg .
Ig (u) =
2 M
2 M
2 M
Let (u ) be a sequence of functions in H12 (M ). Following standard terminology,
we say that (u ) is a Palais-Smale sequence for equation (E ), or that (u ) is a
Palais-Smale sequence for Ig , if the two following propositions hold:
the sequence (Ig (u )) is bounded and
DIg (u ) 0 strongly in H12 (M ) as + .
(2.1.2)
As an elementary remark, it should be noted that if (u ) is a bounded sequence

in H12 (M ) of solutions of (E ), then (u ) is a Palais-Smale sequence for (E ).
15
THE MODEL EQUATIONS
We assume in what follows that the operators L

g are uniformly coercive. More
precisely, we assume that the following holds:
there exists h smooth, s.t. g + h is coercive, and h h for all .
(2.1.3)
Taking inspiration from Brezis-Nirenberg [15], thanks to the mountain pass lemma
of Ambrosetti and Rabinowitz [4], we prove the following proposition in this section.
P ROPOSITION 2.1 Let (M, g) be a smooth compact Riemannian manifold of dimension n 3, and (h ) be a sequence of smooth functions on M . We assume that
(2.1.3) holds. Then there exist Palais-Smale sequences of smooth positive functions
for (E ).
Proof. We let ( ) be a sequence of positive real numbers such that 0 as
+, and let q (1, 2 1). We fix , and let be the functional defined
on H12 (M ) by
1
1
2
|u| dvg +
h u2 dvg
(u) =
2 M
2 M

1

(u+ )2 dvg
(u+ )q+1 dvg
2 M
q+1 M
where u+ = max(u, 0). It is easily seen that the mountain pass lemma of Ambrosetti and Rabinowitz [4], as stated in Brezis and Nirenberg [15], can be applied
to . It follows that there exists c > 0 and a sequence (j ) in H12 (M ) such that
(j ) = c + o(1) and
D (j )(H12 ) = o(1)
(2.1.4)
where o(1) 0 as j +. Moreover, given u0 H12 (M ), u0 0, and u0 0,

we can choose c such that
c sup (tu0 ) .
t0
We claim that by taking for u0 local test functions as in Aubin [6], c can be chosen
such that c < n1 Knn . More precisely, given > 0 small, x0 in M , and > 0,
we let u be the function

1 n2
u (x) = (x) 2 + dg (x0 , x)2
where dg is the distance with respect to g, and is a smooth cutoff function such
that = 1 in Bx0 () and = 0 in M \Bx0 (2), the geodesic balls of center x0 and
radii and 2. Easy computations give that for any q > 1 if n 4, and any q > 3
if n = 3,

u
1
n n2
n n2
2 (q+1) + o
2 (q+1)
=
sup t
u 2
nKnn
t0
where A = A(q, n) is a positive constant depending only on q and n. Assuming
that q > 3 if n = 3, and letting u0 = u with > 0 sufficiently small, this proves
16
CHAPTER 2
the above claim. Following (2.1.4),

1
|j |2 + h 2j dvg
2 M
2
=
(+
)
dv
+
(+ )q+1 dvg + c + o(1)
g
2 M j
q+1 M j
and

|j |2 + h 2j dvg
M
2
q+1
=
(+
)
dv
+
(+
dvg + j H12 o(1) .
g
j
j )
(2.1.5)
(2.1.6)
1
q+1 (2.1.6),
we get that
Writing (2.1.5)

1
1
|j |2 + h 2j dvg c + o(1) + j H12 o(1) .

2 q+1
M
Because of (2.1.3), and since c < nKnn , we have proved that there exists C > 0,
independent of , such that for any j, j H12 C. Up to a subsequence we may
therefore assume that j u weakly in H12 (M ), that j u strongly in
Lp (M ) when p < 2 , and that j u almost everywhere as j +. It
follows that
2
g u + h u = (u+
)
q
+ (u+
) .
Regularity results and the maximum principle then give that u is smooth and that
either u 0 or u > 0 everywhere. Up to another subsequence, we can assume
that j 22 as j +. If u 0 we then get with (2.1.5) and (2.1.6)
1/2
as j + and that nc = . On the other hand, thanks to
that +
j 2
Hebey and Vaugon [47, 48], there exists B > 0 such that
+ 2
2
2
2
+
j 2 Kn j 2 + Bj 2
so that, passing to the limit as j +, we get that 2/2 Kn2 . Noting that this
is in contradiction with the equations nc = and c < n1 Knn , we have proved
that u > 0 everywhere. Hence, (u ) is a sequence of smooth positive functions,
solutions of the equations
2 1
+ uq .
g u + h u = u
Moreover, since j H12 C and j u weakly in H12 (M ), we also have that

for any , u H12 C. Then

2
u dvg +
uq+1 dvg
Ig (u ) =
2 2
2 M
M
is uniformly bounded in . Similarly, for any H12 (M ),
DIg (u ). =
uq dvg ,
M
17
THE MODEL EQUATIONS
and since q < 2 1,
DIg (u ).

u2 q/(2 1) dvg
(2 1)/2
2
C H12
where C > 0 is independent of . In particular, the conditions in (2.1.2) are
satisfied. This proves the proposition.
2
A more classical construction of Palais-Smale sequences involves bubbles as
defined in the following chapter.
2.2 STRONG SOLUTIONS OF MINIMAL ENERGY
We assume in what follows that, for any , the operator L
g is coercive. We discuss
the existence of strong solutions of minimal energy. Given u H12 (M ), we define
the energy E(u) of u by E(u) = u2 . If u is a strong solution of (E ),
Ig (u) =

1
E(u)2 .
n
(n2)/2
where Kn is the sharp constant in the Euclidean

We let also min = Kn
Sobolev inequality 2 Kn 2 , C0 (Rn ). A very standard result is
the following.
P ROPOSITION 2.2 Let (M, g) be a smooth compact Riemannian manifold of dimension n 3 and (h ) be a sequence of smooth functions on M . If, for any ,
L
g = g + h is coercive, and

|u|2 dvg + M h u2 dvg
1
M
inf
< 2

2/2

K
2
uH12 (M )\{0}
n
|u| dvg
M
then (E ) possesses a sequence (u ) of smooth positive solutions such that for any
, E(u ) < min .
Proof. The proof of such an assertion is by now very classical. We fix , and let I
be the functional defined on H12 (M ) by
|u|2 dvg +
h u2 dvg .
I(u) =
M
We let also H be the subset of H12 (M ) defined by

2
|u|2 dvg = 1
H = u H1 (M ) s.t.
M
and
= inf I(u) .
uH
is coercive, 0. By definition, a sequence (ui ) in H12 (M ) is a miniSince

mizing sequence for if
L
g
18
CHAPTER 2
(i) ui H for all i and

(ii) I(ui ) as i +.
We let (ui ) be such a minimizing sequence, and assume that the ui s are nonnegative. Clearly, (ui ) is bounded in H12 (M ). After passing to a subsequence, we may
thus assume that there exists u H12 (M ) such that ui u weakly in H12 (M ),
ui u strongly in L2 (M ), and ui u almost everywhere as i +. In particular, u is nonnegative. Independently, it easily follows from the weak convergence
that
ui 22 = (ui u)22 + u22 + o(1)
for all i, where o(1) 0 as i +. We also have (see, for instance, Brezis-Lieb
[14]) that
ui 22 = ui u22 + u22 + o(1)
for all i, where, as above, o(1) 0 as i +. As shown by Hebey and Vaugon

[47, 48], there exists B > 0 such that, for any i,
ui u22 Kn2 (ui u)22 + Bui u22 .
Since ui H, it follows that

2/2
Kn2 ui 22 u22 + o(1) .
1 u22
Since I(ui ) , and since ui u in L2 (M ), we also have that

2
2
2
2
2
2
2
Kn ui 2 u2 = Kn Kn
|u| dvg +
h u dvg + o(1)
M
Kn2 Kn2 u22 + o(1) .

Hence,

2/2
1 u22
Kn2 1 u22 .
We assumed that Kn2 < 1. Noting that

2/2
,
1 u22 1 u22
this implies that u2 = 1. Then, ui 2 u2 as i +, and since
ui 22 = (ui u)22 + u22 + o(1) ,
we get that ui u strongly in H12 (M ) as i +. In particular, u is a minimizer
for , > 0, and u is a weak nonnegative solution of the equation
g u + h u = u2
By standard regularity results, as developed by Trudinger [74], and using the maximum principle, u is smooth and positive. Noting that u = (n2)/4 u is a solution
2
of (E ), this proves the proposition.
19
THE MODEL EQUATIONS
Let Sg be the scalar curvature of g. Local arguments as in Aubin [6] give that if
n 4, and if there exists x M such that
h (x) <
n2
Sg (x) ,
4(n 1)
then the inequality on the infimum in Proposition 2.2 is satisfied. This provides
examples of h s for which (E ) possesses a sequence (u ) of smooth positive
solutions such that for any , E(u ) < min . Another framework where strong
solutions of minimal energy arise is when discussing the sharp constant problems.
As already mentioned, we know from Hebey and Vaugon [47, 48] that for any
smooth compact Riemannian manifold (M, g) of dimension n 3, there exists
B > 0 such that for any u H12 (M ),
u22 Kn2 u22 + Bu22 .
Let B0 (g) be the smallest B in this inequality. Taking u = 1, it is easily seen that
2/n
B0 (g) Vg
, where Vg is the volume of M with respect to g. We also have
that, for any u H12 (M ),
u22 Kn2 u22 + B0 (g)u22 .
By the definition of B0 (g), if h : M R is such that h < Kn2 B0 (g), then the
inequality on the infimum in Proposition 2.2 is satisfied. On the other hand, it was
proved by Hebey and Vaugon [49] that there are compact manifolds for which the
above sharp Sobolev inequality does not possess (nonzero) extremal functions. For
such manifolds, if (h ) is a sequence of smooth positive functions satisfying that
h < Kn2 B0 (g) for all , and that h Kn2 B0 (g) in C 0 (M ) as +,
then equation (E ) with respect to g possesses a sequence (u ) of smooth positive
solutions such that for any , E(u ) < min , such that E(u ) min as the
parameter +, and such that u 2 0 as +. In particular, the
u s necessarily blow up as +. We refer to Chapter 3 for the definition of
blowing-up sequences of solutions.
2.3 STRONG SOLUTIONS OF HIGH ENERGIES
We discuss the existence of strong positive solutions of arbitrary high energies in
some specific cases. We claim that the following result holds.
P ROPOSITION 2.3 There are smooth compact Riemannian manifolds with the following property: there exists 0 > 0 such that for any 1 > 0 there exist positive
real numbers K1 (1 ) < K2 (1 ) such that, if (h ) is any sequence of real numbers
in (K1 (1 ), K2 (1 )), then (E ) possesses a sequence (u ) of smooth nonconstant
positive solutions such that for any , 1 E(u ) 1 + min .
Typical examples of such manifolds are the unit sphere in odd dimension and the
product of a circle with any compact Riemannian manifold. We prove the above
proposition in the second case.
20
CHAPTER 2
Proof. We let (M, g) be any smooth compact Riemannian manifold of dimension

n 1, and let (S 1 (t), ht ) be the circle in R2 of center 0 and radius t together with
its standard metric. We let Mt = S 1 (t) M and gt = ht + g be the product metric
on Mt . Given m integer, we let Gm be the subgroup of O(2) generated by
ze
2i
m
and regard Gm as acting on Mt by (x, y) ((x), y). Then Mt /Gm = Mt/m .

We let u be a smooth nonconstant function on M , and let ut be the function it
induces on Mt by ut (x, y) = u(y). Then, ut = ut for all Gm . We
know from Hebey and Vaugon [47, 48] that there exists B > 0 such that, for any
u H12 (Mt ),
u22 Kn2 u22 + Bu22
where Kn is the sharp Euclidean constant, and 2 the critical Sobolev exponent.
We let B0 (gt ) be the smallest constant B in the above inequality. Then, for any
u H12 (Mt ),
u22 Kn2 u22 + B0 (gt )u22 .
2/n
Taking u = 1, it is easily seen that B0 (gt ) Vgt

where Vgt is the volume of Mt
2/n
with respect to gt . We claim now that for m sufficiently large B0 (g1/m ) > Vg1/m .
If this is not the case, then, up to a subsequence,
u1/m 22 .
u1/m 22 Kn2 u1/m 22 + Vg2/n
1/m
Noting that for all p,
M1/m
|u1/m |2 dvg1/m =
M1/m
|u1/m |p dvg1/m
1
m
1
=
m
M1
M1
|u1 |2 dvg1 and
|u1 |p dvg1 ,
we get that for any m,

Kn2
u1 22 + Vg2/n
u1 22 .
1
m2/n
This in turn implies that, for any m,
u1 22
(2)2/n Kn2
u22 + Vg2/n u22
m2/n
where Vg is the volume of M with respect to g. Letting m +, we get that
u22
u22 Vg2/n u22 ,

and this is impossible since u is nonconstant. The above claim is proved. In partic2/n
ular, there exists 0 > 0 such that, for any 1 > 0 , B0 (g1/m ) > Vg1/m where

2
2
2
greatest integer not exceeding
m = E(2
min 1 ) + 1, and E(min 1 ) is the

2 2
2
min 1 . Given 1 > 0 , we let m = E(2
1 ) + 1. For real B satisfying

min
2/n
the condition Vg1/m < B < B0 (g1/m ), we let

|u|2 + Kn2 Bu2 dvg1/m
M1/m
B =
inf
.

2/2
uH12 (M1/m )\{0}

2
|u| dvg1/m
M1/m
21
THE MODEL EQUATIONS
Since B < B0 (g1/m ), B < Kn2 . Moreover, B is nondecreasing in B, and it

can be proved with the kind of arguments developed in the preceding section that
lim
BB0 (g1/m )
B = Kn2 .
2
Noting that min = (Kn2 )(n2)/4 , and that m > 2
min 1 , we let
Vg2/n
< T1 (1 ) < B0 (g1/m )
1/m
be such that, for any T1 (1 ) < B < B0 (g1/m ),

1/2

(n2)/4
2
1 E(2
B
.
min 1 ) + 1
We set T2 (1 ) = B0 (g1/m ), and let (h ) be any sequence of real numbers in
(K1 , K2 ) where K1 = Kn2 T1 and K2 = Kn2 T2 . We write that h = Kn2 ,
so that T1 < < T2 . We set = . Since < Kn2 , we get with the kind
of arguments developed in the preceding section that there exists a minimizer u
for . Moreover, this minimizer can be chosen smooth and positive. Clearly, u
is not constant. If this is not the case,
Vg2/n
= Kn2 ,
1/m
2/n
and this is impossible since Kn2 < 1 and Vg1/m > 1. Independently, we can
choose u such that it is a solution of the equation
2 1
g1/m u + h u = u
(n2)/4
=
and such that u 2 = 1. If we let u
+ h u
=
g1/m u
(n2)/4
and
u 2 =
Then
u , then
2 1
u
. We let u
be the function on M1 such that u
/Gm = u
.
2 1
+ h u
= u
g1 u

(n2)/4
1/2
. Noting that 21 + 2min
1 +
and E(
u ) =
u 2 = m1/2
min , it is easily seen from the above construction that
u ) 1 + min .
1 E(
This proves the proposition.
As a remark, the same proof gives that for any sequence (h ) of smooth functions on M , if for any , K1 (1 ) < h < K2 (1 ), then (E ) on M1 = S 1 M
possesses a sequence (u ) of smooth nonconstant positive solutions such that, for
any , 1 E(u ) 1 + min .
Another possible construction, mixing both the ideas of the preceding section
and of this section, is as follows. Given m 1 integer, we let (h ) be a sequence
1
) M for which equation (E ) on M1/m
of smooth functions on M1/m = S 1 ( m
possesses a sequence (u ) of smooth positive solutions. Let : M1 M1/m
22
CHAPTER 2
and u
= h and
be the canonical projection, and let h
be such that h
u
= u . Then, for any ,
2 1
u
+ h
= u
g1 u
on M1 , and E(
u ) = m1/2 E(u ). Let us now suppose that the dimension n of M
is such that n 3, and that g is conformally flat. Then, g1/m is also conformally
flat on M1/m . By conformally flat we mean that, up to conformal changes of
the metric, we get local isometries with the Euclidean space. A conformal metric
to some Riemannian metric g is expressed as f g where f is a smooth and positive
function. When n 4, g is conformally flat if and only if Wg 0, where Wg is
the Weyl curvature tensor of g. Following Hebey-Vaugon [49], there exists gm , a
conformal metric to g1/m on M1/m , such that the sharp Sobolev inequality
gm )u22
u22 Kn2 u22 + B0 (
with respect to gm does not possess (nonzero) extremal functions, where B0 (
gm ) is
defined as in the proof of Proposition 2.3 (see also the remark at the end of section
2.2). We refer to Hebey-Vaugon [49] for a more general statement. Let (h ) be
gm ), and
a sequence of positive real numbers such that for any , h < Kn2 B0 (
such that h Kn2 B0 (
gm ) as +. By the definition of B0 (
gm ), the kind
of arguments of the preceding section give that, for any , there exists u smooth
and positive on M1/m , and (0, Kn2 ), such that
2 1
gm u + h u = u
(n2)/4
and E(u ) = 1. We let u

be given by u
=
u . By the definition of
gm ), and since the above sharp Sobolev inequality does not possess extremal
B0 (
be the smooth
functions, Kn2 and u 2 0 as +. Let u
= u , where is as above, and let gm be
positive function on M1 given by u
the conformal metric to g1 on M1 given by gm = gm . Then, for any ,
2 1
gm u
+ h u
= u
.
(n2)/4
, so that E(
u ) m1/2 min as +,
Moreover, E(
u ) = m1/2
and
u 2 0 as +, where the energy and the L2 -norm in these equations
are given with respect to gm . Summarizing, we have proved that, if (M, g) is a
conformally flat Riemannian manifold of dimension n 3, then, for any m 1
integer, there exists a conformal metric gm to the standard product metric on M1 =
S 1 M , there exists a bounded sequence (h ) of positive real numbers, and there
exists a sequence (u ) of smooth positive functions on M1 , solutions of equations

u ) m1/2 min as +,
(E ) on M1 with respect to gm , and such that E(
and
u 2 0 as +, where the energy and the L2 -norm in these equations
s necessarily blow up as +.
are with respect to gm . In particular, the u
Here again, we refer to Chapter 3 for the definition of blowing-up sequences of
solutions. Independently, note that the same construction can be carried out with
M of order m of
S n , n odd, in place of S 1 M , or with any covering = M
a compact Riemannian manifold of dimension n 4 whose Weyl curvature tensor
vanishes on some open subset of M .
23
THE MODEL EQUATIONS
2.4 THE CASE OF THE SPHERE

Let (S n , h) be the unit n-sphere, n 3. Let Sh be the scalar curvature of h. Then
Sh is constant and Sh = n(n 1). We consider the equation

n(n 2)
h u +
u = u2 1
4
n2
so that h = 4(n1)
Sh . The positive solutions of this equation are known. Let
x0 S n and r be the distance to x0 . Then, for any > 1,
n2

4

1 n2
n(n 2) 2
( 1)
cos r
u =
4
is a solution of the above equation. The energy of u is E(u ) = min . A
possible reference in book form is Hebey [44]. It is easily seen that u 0 in
0
(S n \{x0 }) as 1. On the other hand, u (x0 ) + as 1. Thus the
Cloc
u s develop a singularity at x0 (they blow up at x0 ) as 1. Let x and be
given by
1 n
2
u (x ) = maxn u (x) =
xS
Then x = x0 and
4( 1)
.
n(n 2)( + 1)
In particular, 0 as 1. Let B be the function given by
n2
2
B (x) =
.
d (x ,x)2
2 + hn(n2)
We refer to B (see the following chapters) as the standard bubble with respect to
x and . It is easily seen that there exists C > 1 such that
1
B (x) u (x) CB (x)
(2.4.1)
C
for all x S n and all > 1. Similarly, for any > 0, there exists > 0 such that
1
B (x) u (x) (1 + )B (x)
(2.4.2)
1+
for all x Bx0 ( ) when > 1 is close to 1. Let R be the function given by
u = B + R .
We claim that

R 0 in H12 (S n )
as 1. Noting that S n B2 dvh 0 as 1, we get with (2.4.1) that u 0

in L2 (S n ) as 1, and then that R 0 in L2 (S n ) as 1. Independently
(see, for instance, Chapter 3)

|B |2 dvh 2min and
B2 dvh 2min
Sn
Sn
24
CHAPTER 2
as 1. Moreover, since E(u ) = min and u 0 in L2 (S n ), and thanks to

the equation satisfied by u ,

|u |2 dvh 2min
Sn
as 1. Writing that
2
2
|R | dvh =
|u | dvh +
Sn
Sn
|B | dvh 2
Sn
we then get that

|R |2 dvh = 22min 2
Sn
(u B ) dvh ,
Sn
(u B ) dvh + o(1)
(2.4.3)
Sn
where o(1) 0 as 1. Because of the equation satisfied by u , and since

u 0 and B 0 in L2 (S n ),

(u B ) dvh =
u2 1 B dvh + o(1) .
Sn
Sn
n

Given > 0, u 0 and B 0 in
S \Bx0 () as 1. Writing that

u2 1 B dvh =
u2 1 B dvh +
u2 1 B dvh
0
Cloc
Bx0 ()
Sn
S n \Bx0 ()
we then get with (2.4.2) that

u2 1 B dvh = 2min + o(1) .
Sn
Hence,
Sn
(u B ) dvh = 2min + o(1) ,

and coming back to (2.4.3) we get that
|R |2 dvh 0
Sn
as 1. In particular, R 0 in H12 (S n ) as 1. The above claim is proved.

On the other hand, in general, R 0 in C 0 (S n ) as 1, and the R s are not
even bounded in L (S n ). Let us assume, for instance, that n 7, and let (y ) be
a sequence of points in S n such that dh (x0 , y )2 = ( 1)1+ where > 0 is such
that < n6
4 . Then, R (y ) + as 1. More complete information is
as follows. When n 7, the R s are not bounded in L (S n ). When n = 6, the
R s are bounded in L (S n ), but they do not converge to 0 in C 0 (S n ) as 1.
When n = 3, 4, 5, the R s converge to 0 in C 0 (S n ) as 1. Refinements of
these constructions are in Druet-Hebey [31].
Chapter Three
Blow-up Theory in Sobolev Spaces
An important result of the 1980s describes the asymptotic behavior of Palais-Smale
sequences associated with equations like

u = u2 1 ,
(3.0.1)
where is the Euclidean Laplacian, and u is required to vanish on the boundary
of a smooth bounded open subset of the Euclidean space Rn . This result was
proved by Struwe [73]. Related references are Brezis [11], Brezis-Coron [12, 13],
Lions [58], Sacks-Uhlenbeck [64], Schoen [67], and Wente [76]. Let D12 (Rn ) be
the homogeneous Sobolev space defined as the completion of C0 (Rn ), the space
of smooth functions with compact support in Rn , with respect to the norm

uD12 =
|u|2 dx .
Rn
Nonnegative solutions in D12 (Rn ) of (3.0.1) have been classified by Caffarelli, Gidas, and Spruck [17]. We refer also to Obata [60]. The result we use is as follows:
if u D12 (Rn ), u nonnegative and nontrivial, is a solution of (3.0.1), then u = ua
for some a Rn and some > 0, where
n2

2
.
(3.0.2)
ua (x) =
2
1 + n(n2)
|x a|2
These functions are extremal functions for the sharp Euclidean Sobolev inequality
22

2
2
|u| dx
Kn
|u|2 dx .
(3.0.3)
Rn
Rn
Up to a nonzero multiplicative constant, they are the only nontrivial extremal functions for (3.0.3). Their energy is precisely the minimum energy in the sense that
E(ua ) = ua 2 = min . The free energy Ef , defined for u D12 (Rn ), is given
by

1
1
|u|2 dx
|u|2 dx .
(3.0.4)
Ef (u) =
2 Rn
2 Rn
It is easily checked that if u = ua is a nonnegative nontrivial solution of (3.0.1),
then Ef (u) = n1 Knn .
We present in this chapter the Struwe result [73] for smooth compact Riemannian
manifolds and equations like

(E )
g u + h u = u2 1
as considered in the preceding chapter. For the sake of clarity, we assume that the
h s are uniformly bounded and that they converge L2 to some limiting function.
26
CHAPTER 3
3.1 THE H12 -DECOMPOSITION FOR PALAIS-SMALE SEQUENCES

Given (M, g) smooth, compact, of dimension n 3, we let (h ) be a sequence of
smooth functions on M . We assume that there exists C > 0 and a smooth (or only
continuous) function h on M such that the following holds:
for any and any x, |h (x)| C and
(3.1.1)
h h in L2 (M ) as + .
It clearly follows from (3.1.1) that, for any p 1, h h in Lp (M ) as the parameter +. Let (u ) be a Palais-Smale sequence of nonnegative functions
for (E ). If

1
1
1
|u|2 dvg +
h u2 dvg
|u|2 dvg ,
Ig (u) =
2 M
2 M
2 M
the u s are characterized by the following:
the sequence (Ig (u )) is bounded and
(3.1.2)
DIg (u ) 0 strongly in H12 (M ) as + .
We let Ig be the functional defined on H12 (M ) by

1
1
1
2
2
|u| dvg +
h u dvg
|u|2 dvg
Ig (u) =
2 M
2 M
2 M
and (E ) be the equation

g u + h u = u2 1
(E )
where h is as above. In the following, given > 0, denotes a smooth cutoff
function in Rn such that = 1 in B0 () and = 0 in Rn \B0 (2). For x M ,
where (M, g) is a smooth compact Riemannian manifold, and < ig /2, where ig
is the injectivity radius, we let ,x be the smooth cutoff function in M given by

,x (y) = exp1
x (y)
where expx is the exponential map at x. Here and in what follows, we regard
expx as defined in Rn . An intrinsic definition is possible if M is parallelizable.
i , i = 1, . . . , N , be open subsets of M such that for any
If not we let i and
i , and such that M = i . The canonical
i, i is parallelizable and i
exponential map gives N maps expx defined in i Rn , and expx is, depending
on the situation, one of these maps. A property of expx that holds for any x M
should then be regarded as a property that holds for any i and any x i .
Given a converging sequence (x ) in M , and a sequence (R ) of positive real

numbers such that R + as +, we define a bubble B as a sequence
of functions on M such that
n2

B (x) = (x)R 2 ua R exp1
x (x)
where = ,x , < ig /2, and ua is a nontrivial nonnegative solution of (3.0.1),
thus given by (3.0.2). Roughly speaking, the Struwe result [73] we prove in this
section is that, if (u ) is a Palais-Smale sequence for (E ), then, in the H12 -sense,
u = solution of the limit equation + sum of bubbles
with the additional property that the energies split also. An equivalent definition of
bubbles is in Chapter 6. A more precise statement is as follows:
27
BLOW-UP THEORY IN SOBOLEV SPACES
T HEOREM 3.1 Let (M, g) be a smooth compact Riemannian manifold of dimension n 3, (h ) be a sequence of smooth functions on M such that (3.1.1) is
satisfied and (u ) be a Palais-Smale sequence of nonnegative functions for (E ).
j
j
j
), R
> 0, and R
+ as , conThere exist m N, sequences (R
j
verging sequences (x ) in M , a nonnegative solution u0 H12 (M ) of (E ), and
nontrivial nonnegative solutions uj = uajj D12 (Rn ) of (3.0.1), j = 1, . . . , m,

such that, up to a subsequence,
u = u0 +
m
j uj + o(1)
(3.1.3)
j=1
where

j n2
j
2
uj (x) = R
uj R
exp1
,
j (x)
x
j = ,xj , < ig /2, and o(1)H12 0 as +. Moreover,

Ig (u ) = Ig (u0 ) +
m n
K + o(1)
n n
(3.1.4)
where o(1) 0 as .
The proof of Theorem 3.1 proceeds in several steps. We follow the proof of
Struwe [73]. Uniqueness of the decomposition (3.1.3) is, for instance, discussed
by Hebey [46]. We use the terminology of a Palais-Smale sequence for (Ig ) instead
of a Palais-Smale sequence for (E ). First we claim that the following result holds:
S TEP 1. Palais-Smale sequences for (Ig ) are bounded in H12 (M ).
Proof of step 1. We let (u ) be a Palais-Smale sequence for (Ig ). Then,

|u |2 dvg +
h u2 dvg
|u |2 dvg
DIg (u ).u =
M
M
M

= o u H12 .
Hence,
Ig (u )
1
=
n

|u |2 dvg + o u H12
and, since Ig (u ) C for some C > 0 independent of , we get that

|u |2 dvg C + o u H12 .
M
Thanks to Holders inequality, this implies in turn that

2/2
u2 dvg C + o u H 2
M
where C > 0, like all the constants below, is independent of . Writing that

2
|u |2 dvg +
h u2 dvg = 2Ig (u ) +
|u |2 dvg ,
2 M
M
M
28
CHAPTER 3
we also get that
|u |2 dvg +

h u2 dvg C + o u H12 .
Noting that, thanks to (3.1.1),

|u |2 + h u2 dvg + Cu 22 ,
u 2H 2
1
it follows from the above equations that

2/2
u 2H 2 C + o u H12 + o u H 2
.
1
This clearly implies that the u s are bounded in
H12 (M ).
Step 1 is proved.
A second step in the proof is as follows. We let Ig be the functional defined on

H12 (M ) by

1
1
|u|2 dvg
|u|2 dvg
Ig (u) =
2 M
2 M
so that Ig = Ig when h 0.
S TEP 2. Let (u ) be a Palais-Smale sequence of nonnegative functions for Ig ,
and u0 H12 (M ) a nonnegative function such that u u0 weakly in H12 (M ),
u u0 strongly in L2 (M ), and u u0 almost everywhere as +. Let
u ) is a Palais-Smale sequence for Ig and
u
= u u0 . Then (
Ig (
u ) = Ig (u ) Ig (u0 ) + o(1)
where o(1) 0 as +. Moreover, u0 is a solution of (E ).
Proof of step 2. We first observe that for any C (M ),
2 1
(u ) dvg +
h u dvg
u
dvg
DIg (u ). =
M
(3.1.5)
= o(1)
where o(1) 0 as +, and (u ) is the scalar product with respect to
g of u and . By (3.1.1), h h in Lp (M ) for any p 1 as +.
Writing that
h u dvg =
(h h )u dvg +
h u dvg
(3.1.6)
M
and noting that, from Holders inequalities,

(h h )u dvg u 2 h h n

2
1/2
|| dvg
we easily get by passing to the limit in (3.1.5) that

0

h u0 dvg =
(u0 )2 1 dvg .
u dvg +
M
(3.1.7)
29
. We know by
Hence, u0 is a solution of (E ). Now we compute the energy of u
step 1 that the u s are bounded in H12 (M ). From the Sobolev embedding theorem
and (3.1.6), (3.1.7) with = u , we can thus write that
h u2 dvg =
h (u0 )2 dvg + o(1) .
M
It follows that
Ig (u )
Ig (u0 )
1
+ Ig (
u )
2
dvg + o(1)
(3.1.8)
u +u0 |2 |
u |2 |u0 |2 . Clearly, there exists C > 0, independent
where = |
of , such that

| | dvg C
|
u |2 1 |u0 |dvg + C
|u0 |2 1 |
u |dvg
while basic integration theory gives that

|
u |2 1 |u0 |dvg = o(1) and
M
Therefore,

M
|u0 |2
|
u |dvg = o(1) .
dvg = o(1), and coming back to (3.1.8), we get that

Ig (
u ) = Ig (u ) Ig (u0 ) + o(1) .
Summarizing, we are left in the proof of step 2 with the proof that (
u ) is a PalaisSmale sequence for Ig . Let C (M ). From (3.1.6), (3.1.7), and the Sobolev
embedding theorem,

h u dvg =
h u0 dvg + o H12 .
M
It follows that
DIg (u ).
= DIg (
u ).
M

dvg + o H12
(3.1.9)
where
u + u0 |2
= |
(
u + u0 ) |
u |2
u
|u0 |2
2 0
u .
Here again, it is easily checked that there exists C > 0, independent of , such that

| |dvg C
|
u |2 2 |u0 |||dvg + C
|u0 |2 2 |
u |||dvg .
M
From Holders inequalities we then get that
| |dvg
M

0 2 2

C |
u |2 2 u0
+ |
u |
u

2 /(2 1)
2 /(2 1)
while basic integration theory gives that

0 2 2

= o(1) and |
u |
u

u |2 2 u0
|
2 /(2 1)
2 /(2 1)
2
= o(1) .
30
CHAPTER 3
Coming back to (3.1.9), and from the Sobolev embedding theorem, we then get
that

DIg (u ). = DIg (
u ). + o H12 .
This implies that (
u ) is a Palais-Smale sequence for Ig . Step 2 is proved.
A third step in the proof of Theorem 3.1 is the following. We let =

be the value of the free energy Ef when considered on the ua s of (3.0.2).
1
n
n Kn
u ) be a Palais-Smale sequence for Ig such that u

0 weakly
S TEP 3. Let (
u ) as +, where < .
in H12 (M ) as +, and such that Ig (
Then u
0 strongly in H12 (M ).
Proof of step 3. By step 1, the u
s are bounded in H12 (M ). Independently,

2
u ).
u =
|
u | dvg
|
u |2 dvg
DIg (
M
M

= o
u H12 .
Hence,

1
1

u 22 + o(1) =
u 22 + o(1) = + o(1) .
(3.1.10)
n
n
This already implies that 0. Following Hebey-Vaugon [47, 48], there exists
B > 0, independent of , such that
Ig (
u ) =

u 22 Kn2
u 22 + B
u 22 .
Moreover, since the embedding H12 (M ) L2 (M ) is compact, u
0 strongly
in L2 (M ) as +. Letting + in the above sharp Sobolev inequality,
it follows from (3.1.10) that
(n)2/2 Kn2 n .
Since < , this implies that = 0. By (3.1.10) we then get that u

0
2
strongly in H12 (M ). This proves step 3.
It follows from steps 13 that if (u ) is a Palais-Smale sequence for Ig , and
as +, where < , then, up to a subsequence, (u )
converges strongly to some u0 in H12 (M ). In other words, compactness holds for
Palais-Smale sequences when the energy is below the minimum energy. Another
illustration of this fact is given by Proposition 2.2.
Ig (u )
The following step is the main ingredient in the proof of Theorem 3.1. We postpone its proof to section 3.2.
S TEP 4. Let (
0 weakly in
H12 (M ) but not strongly. Then there exists a sequence (R ) of positive real numbers, R + as , a converging sequence (x ) in M , and a nontrivial
solution u D12 (Rn ) of the Euclidean equation
u = |u|2
(3.1.11)
31
such that, up to a subsequence, the following holds: if

v = u
B
where
n2

B (x) = R 2 u R exp1
x (x)
and = ,x , < ig /2, then (

v ) is also a Palais-Smale sequence for Ig ,
v ) = Ig (
u ) Ef (u) + o(1),
v 0 weakly in H12 (M ) as +, and Ig (
where o(1) 0 as +.
Following steps 14, we are now in a position to prove Theorem 3.1. The proof
proceeds as follows.
Proof of Theorem 3.1. A preliminary claim is that nontrivial solutions of the Euclidean equation (3.1.11) have a free energy Ef bounded from below by . Indeed, if u D12 (Rn ) is a nontrivial solution of (3.1.11), we get with the sharp
Euclidean Sobolev inequality (3.0.3) that

2 /2

|u|2 dx =
|u|2 dx Kn2
|u|2 dx
.
Hence,

Rn
Rn
Rn
|u| dx
Rn
Knn
and

1
1

|u|2 dx .
Ef (u) =
2 2
Rn
This proves the above claim. In order to prove the theorem, we let (u ) be a PalaisSmale sequence of nonnegative functions for Ig . According to step 1, (u ) is
bounded in H12 (M ). Up to a subsequence, we may therefore assume that for some
u0 H12 (M ), u u0 weakly in H12 (M ), u u0 strongly in L2 (M ), and
u u0 almost everywhere as +. We may also assume that Ig (u ) c
as +. By step 2, u0 is a nonnegative solution of (3.0.1) and u
= u u 0
is a Palais-Smale sequence for Ig such that
Ig (
u ) = Ig (u ) Ig (u0 ) + o(1) .
If u
0 strongly in H12 (M ), then u = u0 + o(1), and the theorem is proved. If
not, according to the claim at the beginning of this proof, we apply step 4 to get a
new Palais-Smale sequence (
u1 ) of energy
Ig (
u1 ) Ig (
u ) + o(1) .
Here again, either u
1 0 strongly in H12 (M ), in which case the theorem is proved,
1
or u
0 weakly but not strongly in H12 (M ), in which case we again apply
step 4. By induction, at some point, either we do have compactness or the PalaisSmale sequence (
um
) we get with this process has an energy that converges to

2
m
some < . Then, by step 3, u
0 strongly in H1 (M ). Applying step 4
j
in this process, we got some u s, j = 1, . . . , m. Assuming that the uj s are also
nonnegative (see below) this ends the proof of Theorem 3.1.
2
The fact that the uj s we get in the above process are nonnegative is proved in
the next section. We refer to Lemma 3.3.
32
CHAPTER 3
3.2 SUBTRACTING A BUBBLE AND NONNEGATIVE SOLUTIONS

In this section, we prove step 4 and we prove that the uj s in Theorem 3.1 are
nonnegative. We start with the proof of step 4. The notations are those of section
3.1.
L EMMA 3.2 Let (
0 weakly
in H12 (M ) but not strongly. Then there exist a sequence (R ) of positive real numbers, R + as , a converging sequence (x ) in M , and a nontrivial
solution u D12 (Rn ) of the Euclidean equation
u = |u|2
such that, up to a subsequence, the following holds: if

v = u
B
where
n2

B (x) = R 2 u R exp1
x (x)
and = ,x , < ig /2, then (

v ) is also a Palais-Smale sequence for Ig ,
v ) = Ig (
u ) Ef (u) + o(1),
v 0 weakly in H12 (M ) as +, and Ig (
where o(1) 0 as +.
u ) as +. We
Proof. Up to a subsequence, we may assume that Ig (
may also assume that u
is smooth, since if not there always exists u smooth and
H12 0. Then, (u ) is a Palais-Smale sequence for Ig such
such that u u
that u 0 weakly in H12 (M ) but not strongly, and, as is easily checked, if the
u ). Since DIg (
u ) 0, we get as in
claim holds for (u ), then it holds also for (
step 3 of section 3.1 that
|
u |2 dvg = n + o(1)
(3.2.1)
M
and that n Knn . For t > 0, we let

(t) = max
xM
Bx (t)
|
u |2 dvg .
Given t0 > 0 small, it follows from (3.2.1) that there exist x0 M and 0 > 0
such that, up to a subsequence,
|
u |2 dvg 0
Bx0 (t0 )
for all . Then, since t (t) is continuous, we get that for any (0, 0 ),
there exists t (0, t0 ) such that (t ) = . Clearly, there also exists x M
such that
(t ) =
Bx (t )
|
u |2 dvg .
Up to a subsequence, (x ) converges. We let r0 (0, ig /2) be such that for all

x M and all y, z Rn , if |y| r0 and |z| r0 , then
dg (expx (y), expx (z)) C0 |z y|
33
for some C0 [1, 2] independent of x, y, and z. Given R 1 and x Rn such

that |x| < ig R , we let

n2
1
expx (R
x) ,
u
(x) = R 2 u

1
x) .
g (x) = expx g (R
Then,

1
n
x) = R
|
u |2 (x) ,
|
u |2 expx (R
where the norm in the left hand side of this equation is with respect to g, and the
norm in the right hand side is with respect to g . It follows that, if |z| + r < ig R ,
then
2
|
u | dvg =
(3.2.2)
|
u |2 dvg .
1
Bz (r)
expx (R
Bz (r))
When |z| + r < r0 R ,

1
Bz (r) Bexpx
expx R
while
1
(R
z)

1
C0 rR

1

1
expx R
B0 (C0 r) = Bx C0 rR
.
(3.2.3)
(3.2.4)
C0 rt1
0
1. Then, for any (0, 0 ), to

Given r (0, r0 ), we fix t0 such that
1
= t . By (3.2.2)(3.2.4),
be fixed later on, we let R 1 be such that C0 rR
n
for any z R such that |z| < r0 R r,
|
u |2 dvg and
Bz (r)
(3.2.5)
|
u |2 dvg = .
B0 (C0 r)
We let (0, ig ) and C1 > 1 be such that for any x M , and any R 1, if
gx,R (y) = expx g(R1 y), then
1
2
2
|u| dx
|u| dvgx,R C1
|u|2 dx
(3.2.6)
C1 Rn
Rn
Rn
for all u D12 (Rn ) such that suppu B0 (R). Without loss of generality, we also
assume that
1
|u|dx
|u|dvgx,R C1
|u|dx
(3.2.7)
C1 Rn
Rn
Rn
for all u L1 (Rn ) such that suppu B0 (R). We let C0 (Rn ) be a cutoff
function such that
1, = 1 in B0 (1/4), and = 0 in Rn \B0 (3/4). We
10
1
set (x) = R x , where is as above. Then,
|(
u
)|2 dvg = O(1)
Rn
) is bounded in D12 (Rn ). In

and it follows from (3.2.6) that the sequence (
u
u
particular, up to a subsequence, there exists u D12 (Rn ) such that u
34
CHAPTER 3
weakly in D12 (Rn ). Now we divide the proof of Lemma 3.2 into several steps. As
a first step, we claim that the following holds.
S TEP 1. For r and sufficientl small,
u
u strongly in H12 (B0 (C0 r))
(3.2.8)
as +.
Proof of step 1. We let x0 Rn , and for > 0, we let h be the standard metric
on Bx0 (). By Fatous lemma,

2r
N (
u
)dvh d lim inf
N (
u
)dx C
lim inf
r
Bx0 ()
2
Bx0 (2r)
where Nh (u) = |u| + u , the norm in Nh is with respect to h, and is the Euclidean metric. It follows that there exists [r, 2r] such that, up to a subsequence,
and for all ,
N (
u
)dvh C .
Bx0 ()
As an easy consequence, we get that

u
H 2 (Bx
)C
where C > 0 is independent of . The embedding
1
0 ()
2
H12 (Bx0 ()) H1/2
(Bx0 ())
is compact, and the trace operator u u|B is continuous. It follows that, up to a

subsequence,
2
u
u in H1/2
(Bx0 ())
as +. Let A be the annulus A = Bx0 (3r)\Bx0 (). Let also D12 (Rn )
u in Bx0 ( + ), and = 0 in Rn \Bx0 (3r ), > 0
be such that = u
small. Then

u
uH 2 (Bx ()) = H 2 (Bx ())
0
0
1/2
1/2
while there exists 0 D12 (A), the closure of C0 (A) in H12 (A), such that
2 (A) .
+ 0 H12 (A) C H1/2
Minimization arguments give that there exists z H12 (A) such that
z = 0 in A ,
z 0 D12 (A) ,
and
z H12 (A) C + 0 H12 (A) .
Hence, z 0 strongly in H12 (A) as +. We let D12 (Rn ) be such
that
= u
u in B x0 () ,
= z in B x0 (3r)\Bx0 () ,
35
and = 0 otherwise. We let r be such that r < min(ig /6, /24), and let be
such that
n2

(x) = R 2 R exp1
x (x)

if dg (x , x) < 6r, and = 0 otherwise. Clearly, 1 exp1
x (x) = 1 if
dg (x , x) < 6r. If in addition |x0 | < 3r, then
DIg (
u ). = DIg (
u
).
=
((
u
) ) dvg
Bx0 (3r)
Bx0 (3r)
|
u
|2
(
u
) dvg

where (x) = 1 exp1

x (x) . We have that H12 (M ) C D12 (Rn ) . In
u ). = o(1).
particular, the s are bounded in H12 (M ). It follows that DIg (
2
Noting that 0 strongly in H1 (A), and 0 weakly in D12 (Rn ),
((
u
) ) dvg
Bx0 (3r)
Bx0 ()
Rn
(( + u) ) dvg + o(1)
| |2 dvg + o(1) .
Similarly, one easily gets that

|
u
|2 2 (
u
) dvg =
Bx0 (3r)
| |2 dvg + o(1) ,
Rn
and since DIg (

u ). = o(1), we have proved that

2
| | dvg
| |2 dvg = o(1) .
Rn
Rn
(3.2.9)
By the strong convergence 0 in H12 (A), and the weak convergence 0

in D12 (Rn ),
| |2 dvg =
|(
u
u)|2 dvg + o(1)
Rn
Bx0 ()
=
so that
Bx0 ()
Rn
|(
u
)| dvg
| | dvg
Bx0 ()
Bx0 ()
|u|2 dvg + o(1)
|(
u
)|2 dvg + o(1) .
Let N be an integer such that B0 (2) is covered by N balls of radius 1 and centered
in B0 (2). Then there exist N points xi Bx0 (2r), i = 1, . . . , N , such that
Bx0 () Bx0 (2r)
N

i=1
Bxi (r)
36
CHAPTER 3
and we get with (3.2.5) that for x0 and r such that |x0 | + 3r < r0 ,
| |2 dvg N + o(1) .
Rn
(3.2.10)
Independently, from the Sobolev inequality, for C1 as in (3.2.6) and (3.2.7), and
x0 and r such that |x0 | + 3r < , we also have that

2/2

2/2
2/2
2
2
| | dvg
C1
| | dx
Rn
Rn
2/2
C1 Kn2
| |2 dx
Rn
1+(2/2 ) 2
C1
Kn
| |2 dvg .
Rn
Following (3.2.9) and (3.2.10), we can then write that
| |2 dvg K
| |2 dvg + o(1)
Rn
Rn
where
1+(2 /2)
K = C1
Let > 0 be such that

1+(2 /2) 2
Kn (N )(2 /2)1
C1
so that 0 strongly in
(2 /2)1
Kn2 (N + o(1))
< 1. Then,
| |2 dvg = o(1)
Rn
D12 (Rn )
as +. Since r , it follows that
u strongly in H12 (Bx0 (r))

u
(3.2.11)

1+(2 /2) 2
Kn (N )(2 /2)1
C1
< 1, |x0 | < 3r,

and the convergence holds as soon as
|x0 | + 3r < r0 , |x0 | + 3r < , and r < min(ig /6, /24). We fix > 0 sufficiently

1+(2 /2) 2
Kn (N )(2 /2)1 < 1, and let r > 0 be sufficiently
small such that C1
small such that r < min(ig /6, /24, r0 /6). Then (3.2.11) holds for any x0 such
that |x0 | < 2r. Since C0 2, B0 (C0 r) is covered by N balls of radius r and
u strongly in H12 (B0 (C0 r)). This
centered in B0 (2r). It follows that u
proves (3.2.8).
2
From (3.2.5) and (3.2.8), we can write that
|
u |2 dvg
=
B0 (C0 r)
=
|(
u
)|2 dvg
B0 (C0 r)
C1
|u|2 dx + o(1) .
B0 (C0 r)
It follows that u 0. Let us assume that R R as +, R 1. If

0 weakly in H12 (M ).
R < +, then u
0 weakly in H12 (B0 (C0 r)) since u
From (3.2.8), and since u 0, we get that
lim R = + .
(3.2.12)
37
We claim now that the following holds.

S TEP 2. For any R > 0,
u
u strongly in H12 (B0 (R))
as +, and u is a solution of the Euclidean equation u = |u|2
(3.2.13)
u.
Proof of step 2. We let R 1 be given. By (3.2.12), R R for large, and

(3.2.5) holds for z such that |z| < r0 R r. Then, as is easily checked from
the proof of step 1, (3.2.11) holds if |x0 | < 3r(2R 1), |x0 | + 3r < r0 R, and
u
|x0 | + 3r < R. In particular, (3.2.11) holds if |x0 | < 2rR. Hence, u
strongly in H12 (B0 (2rR)). Noting that for x in a compact subset of Rn , (x) = 1
for large, and that R 1 is arbitrary, we easily get that (3.2.13) holds. Now we

prove that u is a solution of the critical Euclidean equation u = |u|2 2 u. Let
C0 (Rn ) and R0 > 0 be such that supp B0 (R0 ). Let also be given by
n2
(x) = R 2 (R x) .
1
R0 ). For large, we let be the smooth function on M
Then supp B0 (R
given by = expx . For large,
(
u ) dvg =
((
u
)) dvg
Rn
and
|
u |2
u
dvg =
Rn
|
u
|2
(
u
)dvg .
From (3.2.12), g in C 1 (B0 (R)) for any R > 0. Moreover, ( ) is bounded

in H12 (M ). Since (
u ) is a Palais-Smale sequence for Ig , and u
u in
D12 (Rn ), we get by passing to the limit as + in the above equations that

(u) dx =
|u|2 2 udx .
Rn
In other words, u
Rn
D12 (Rn )
is such that u = |u|2
u. This proves step 2.
For x M and (0, /8), we let V be given by

n2

V (x) = (x)R 2 u R exp1
x (x)
(3.2.14)
V and claim that the following holds.

where = ,x
. We let w = u
S TEP 3. The following relations hold. On the one hand,
w 0 weakly in H12 (M )
(3.2.15)
as +. On the other hand,

DIg (V ) 0 and DIg (w ) 0
(3.2.16)
strongly as +. Finally,
u ) Ef (u) + o(1)
Ig (w ) = Ig (
(3.2.17)
38
CHAPTER 3
where o(1) 0 as +
Proof of step 3. We start with the proof of (3.2.15). It suffices to prove that V 0
1
R). For a smooth
weakly in H12 (M ). Given R > 0, we let (R) = Bx (R
function on M and large,
n2
V dvg = R 2
(x)u(R x)(expx (x))dvg
1
B0 (R
R)
(R)
where g = expx g. It follows that for C > 0 such that dvg Cdx,

n+2

|u|dx .
V dvg C R 2

(R)

B0 (R)
Similarly, by Holders inequality,

n+2

2
V dvg C R
|u|dx

M \ (R)
B0 (R )\B0 (R)
1/2

C
|u|2 dx
B0 (R )\B0 (R)

Taking R > 0 sufficiently large and using (3.2.12),
we get that M V dvg 0

as +. Similar arguments give that M (V )dvg 0 as +.
This proves (3.2.15). Now we prove (3.2.16). Here again we let be a smooth
function on M . Then,

DIg (V ). =
(V )dvg
|V |2 2 V dvg .
M
Given R > 0, we write that
(V )dvg =
(R)
(V )dvg +
Easy computations give
Bx ()\ (R)
Bx ()\ (R)
(V )dvg .

(V )dvg = O H12 R
where R 0 as R +. Let be the function of D12 (Rn ) given by

1
n2
(x) = R 2 ,(x) expx (R
x)
1
x). Then, for large,
where ,(x) = (R
(V )dvg =
(u )dvg .
(R)
B0 (R)
Noting that g in C 1 (B0 (R )), R > R, and that
2
|| dvg =
| |2 dvg ,
(R)
B0 (R)
39
we get that
B0 (R)
(u )dvg =
We also have that
B0 (R)
B0 (R)
(u )dx =
Rn

(u )dx + o H12 .

(u )dx + O H12 R
where R is as above. Therefore,

(V )dvg =
(u )dx+o H12 +O H12 R . (3.2.18)
Rn
In a similar way, we can prove that

2 2
|V |
V dvg =
|u|2 2 u dx
n
M

R
+ o H12 + O H12 R .
Since u is a solution of u = |u|2
(3.2.19)
u, it follows from (3.2.18) and (3.2.19) that

DIg (V ). = o H12 + O H12 R .

Since R > 0 is arbitrary, we get that DIg (V ) 0 strongly as +. This

proves the first claim of (3.2.16). Now we write
DIg (w ). = DIg (
u ). DIg (V ). A(). ,
where
(3.2.20)
A(). =
dvg =
M
Bx (2)
dvg
and
= |w |2
w |
u |2
u
+ |V |2
V .
By the Holder and Sobolev inequalities,

|A().| 2 /(2 1) H12 .
Given R > 0, we let (R)c = Bx (2)\

(R), where (R) is as above. Then,
for large,
2 /(2 1) L2 /(2 1) ( (R)) + L2 /(2 1) ( (R)c ) .
As in the proof of step 2 of section 3.1, we can write
L2 /(2 1) ( (R)c )

C 1 L2 /(2 1) ( (R)c ) + 2 L2 /(2 1) ( (R)c )
u |2 2 V and 2 = |V |2 2 u
. We have that
where 1 = |

2
| 221 dvg
| | 2 1 dvg =
|
(R)
B0 (R)
40
CHAPTER 3
where

= |
u u|2 2 (
u u) |
u |2 2 u
+ |u|2 2 u .
Using (3.2.13), we then get that
2
(R)
| | 2 1 dvg = o(1) .
Independently,
2
|1 | 2 1 dvg =
)\B0 (R)
B0 (2R
(R)c
|
u
|
C
where
Rn \B0 (R)
|
u
|
2 (2 2)
2 1
2 (2 2)
2 1
2
2
1
|u| 2 1 2
dvg
2
|u| 2 1 dx ,

1
(x) = ,x
expx (R x)
and C > 0 is such that dvg Cdx. Without loss of generality, we may assume
u almost everywhere in Rn . We let
that u
u
|
f = |
2 (2 2)
2 1
and f = |u|
2 (2 2)
2 1
2 1
2 2
Then (f ) is bounded in L
(Rn ) and (f ) converges almost everywhere to f .
From standard integration theory, it follows that
2 1
f f weakly in L 2 2 (Rn )
as +. Hence,
lim
+
Rn \B
0 (R)
|
u
|
2 (2 2)
2 1
and we get that
|u|
lim lim sup
R+ +
Similarly, we can prove that

lim lim sup
R+ +
(R)c
(R)c
2
2 1
dx =
Rn \B
|u|2 dx ,
0 (R)
2
|1 | 2 1 dvg = 0 .
2
|2 | 2 1 dvg = 0 .
Coming back to (3.2.20), and since R > 0 is arbitrary, we get that DIg (w ) 0
strongly as +. This proves (3.2.16). Now we are left with the proof of
(3.2.17). We have that

1
1
Ig (w ) =
|w |2 dvg
|w |2 dvg .
(3.2.21)
2 M
2 M
Concerning the first term in (3.2.21), we write
2
2
|w | dvg =
|w | dvg +
M
Bx (2)
M \Bx (2)
|
u |2 dvg .
41
Then, given R > 0,
|w |2 dvg =
Bx (2)
(R)
|w |2 dvg +
(R)c
|w |2 dvg
where (R) and (R)c are as above. We have
2
|w |2 dvg =
|(
u u)| dvg
(R)
B0 (R)
so that, from (3.2.13),
(R)
|w |2 dvg = o(1) .
Independently, it follows from rough estimates that
lim lim sup

|V |2 dvg = 0 .
R+ +
(R)c
Since w = u
V , and (
u ) is bounded in H12 (M ), we have that
|w |2 dvg =
|
u |2 dvg + BR ()
(R)c
(R)c
where
lim lim sup BR () = 0 .
R+ +
Therefore,
|w |2 dvg =
M
|
u |2 dvg
(R)
where BR () satisfies (3.2.22). Noting that
2
|
u | dvg =
(R)
(3.2.22)
|
u |2 dvg + BR () + o(1)
B0 (R)
|
u |2 dvg
and that g in C 1 (B0 (R)), we get with (3.2.13) that
|
u |2 dvg =
|u|2 dx + o(1)
(R)
B0 (R)
=
|u|2 dx + R + o(1)
Rn
where R 0 as R +. Hence,
2
2
|w | dvg =
|
u | dvg
M
Rn
|u|2 dx + BR () + o(1)
where BR () satisfies (3.2.22). Similarly, we can prove that

|w |2 dvg =
|
u |2 dvg
|u|2 dx + BR () + o(1)
M
Rn
(3.2.23)
(3.2.24)
where BR () satisfies (3.2.22). Combining (3.2.21), (3.2.23), and (3.2.24), we

then get that
Ig (w ) = Ig (
u ) Ef (u) + BR () + o(1) ,
42
CHAPTER 3
and since R > 0 is arbitrary, it follows that

u ) Ef (u) + o(1) .
Ig (w ) = Ig (
This proves (3.2.17) and step 3.
From steps 13, Lemma 3.2 holds for some (0, ig /2) small. Given 1 < 2
in (0, ig /2), it is easily seen that
(2 ,x 1 ,x )B H 2 = o(1) .
1
It follows that Lemma 3.2 holds for any (0, ig /2). This ends the proof of
Lemma 3.2.
2
Another important ingredient we used in the proof of Theorem 3.1 is the following result.
L EMMA 3.3 If (u ) is a Palais-Smale sequence of nonnegative functions for (Ig ),
the uj s of Theorem 3.1 are also nonnegative. Therefore, when j 1, uj = uajj

for some aj Rn and j > 0, where ua is as in (3.0.2).
= u u0 and let
Proof. That u0 is nonnegative is straightforward. We let u
i
i
= 1/R . First we prove the following: for any N integer in [1, m], and for
any s integer in [0, N 1], there exist an integer p and sequences (yj ) and (j ),
j
N
j = 1, . . . , p, yj M , and j > 0, such that for any j, dg (xN
, y )/ is bounded
j
N

and / 0, and such that for any R, R > 0,
2

s

(3.2.25)
ui uN

dvg = o(1) + (R )
u
p
j

N
(R)\j=1 (R )
i=1
j

j
j (R ), (R ) 0 as R +,
where N
(RN
(R) = BxN
), (R ) = By
and the (ui )s and (xi )s are the ordered sequences in i we get in the proof of
Theorem 3.1. We proceed here by inverse induction on s. If s = N 1, then, by
(3.2.13),
2

N
1

ui uN

u
dvg = o(1)

N
(R)
i=1
so that (3.2.25) holds with p = 0. Now we suppose that (3.2.25) holds for some s,
s N 1. If the dg (xs , xN
)s do not converge to 0, then, up to a subsequence,
s
> 0. As a consequence,
(R)
(
R)
=
for
R
N

|us |2 dvg
|us |2 dvg
p

j
N
(R)\j=1 (R )
M \s (R)
and it follows (we refer to the proof of Lemma 3.2) that

|us |2 dvg C
p

j
N
(R)\j=1 (R )
Rn \B
0 (R)
|us |2 dx .
> 0 is arbitrary and us L2 (Rn ), we get that

Since R

|us |2 dvg = o(1)
p

j
N
(R)\j=1 (R )
43
and thus

2
s1

ui uN

u
dvg = o(1) + (R ) .
p
j

N
(R)\j=1 (R )
i=1
In particular, (3.2.25) holds for s 1. Let us now assume that dg (xs , xN

) 0 as
+. We let r0 and C 1 be such that for all x M , and all y, z Rn , if
|y| r0 and |z| r0 , then
1
|z y| dg (expx (y), expx (z)) C|z y| .
C
If x
s and yj are such that xs = expxN
(N
s ) and yj = expxN
(N
j ), then
x
y
j

R
j
1
1 j

j
exp
(R
)
B
C
Byj
R
(3.2.26)
y
xN
C N
N
N
and

Bxs
R s
C N
1
exp1
(s (R )) Bxs
xN
R C
s
N

.
(3.2.27)
> 0, we have by (3.2.13) that

Given R

2
s

i
u dvg = o(1) .

u

s (R)
i=1
Hence, by (3.2.25),
2

|uN
| dvg = o(1) + (R ) ,
p

j
N
(R)\j=1 (R )
s (R)
and it follows from (3.2.26) and (3.2.27) that

|uN |2 dx = o(1) + (R ) .
j
s
B0 (R)\p
j=1 B
j (R
y
C
N
) Bx
(R
s
C
(3.2.28)
Now we distinguish two cases. In the first case we assume that, as +,

N
s
N
s
dg (xs , xN
)/ +. Then we also have that dg (x , x )/ +, since if
s
N
large enough that / 0, while
not, we get by (3.2.28) with R
dg (xs , xN
dg (xs , xN
N
)
)
=
s .
s
N
Then it follows that N

(R) (R) = for R > 0, and we may proceed as in
s
N
the case where the dg (x , x )s do not converge to 0 to get that (3.2.25) holds for
N
s 1. In the second case, we assume that, as +, the dg (xs , xN
)/ s
s
N
p+1
s
converge. By (3.2.28), we must have that / 0. We set y = x and
p+1
= s . Then,

2
s

i
N
u u dvg = o(1) + (R )

u
p+1 j

N
(R)\j=1 (R )
i=1
44
CHAPTER 3
and
|us |2 dvg
p+1 j

N
(R)\j=1 (R )
|us |2 dvg
M \s (R )
(R ) .
It follows that

2
s1

ui uN

u
dvg = o(1) + (R )
p+1 j

N

(R)\j=1 (R )
i=1
and (3.2.25) holds for s 1. In particular, (3.2.25) is always true. Now we prove
the claim that if the u s in Theorem 3.1 are nonnegative, then u0 and the ui s
of Theorem 3.1 are also nonnegative. As already mentioned, it is clear that u0 is
nonnegative. We let u
N
be given by

N n2
N
2 u
u
N
expxN
(
x)
.
(x) = ( )
We apply (3.2.25) with s = 0. Then,

2
u
dvg = o(1) + (R ) ,
uN
p

j
N
(R)\j=1 (R )
and it follows
that

B0 (R)\p
j (R C N
j=1 By
N
2
u
uN dx = o(1) + (R )
(3.2.29)
where the yj are as above. The yj s are bounded. Up to a subsequence we may

assume that yj y j as +. Then we get from (3.2.29) that

N
u
N
in L2loc B0 (R)\X
u
consists of the yj s, j = 1, . . . , p. Therefore, we may
as +, where X
N
N
assume that u
u almost everywhere in Rn as +. Let

N n2
2 u0
u
0,N
(N
expxN
(x) = ( )
x) .
Then,
where g (x) =
|u0 |2 dvg =
N
(R)

expxN g (N
x).
as +, and thus that

vN s given by
u
0,N
u
0,N
2
|
u0,N
| dvg
B0 (R)
It follows that
0 in L2 (B0 (R))
0 almost everywhere in Rn . Summarizing, the
vN (x) = (N
)
n2
2

N
u expxN
(
x)
converge almost everywhere to uN . In particular, uN is nonnegative, and this

proves Lemma 3.3.
2
As a remark, it follows from the above proof that for any i = j,
j
R
Ri
i j
+
+ R
R dg (xi , xj )2 +
j
i
R
R
as +.
45
3.3 THE DE GIORGINASHMOSER ITERATIVE SCHEME FOR

STRONG SOLUTIONS
Given (M, g) smooth, compact, of dimension n 3, we let (h ) be a sequence
of smooth functions on M . We also let (u ) be a bounded sequence in H12 (M ) of
nonnegative solutions of
2 1
.
g u + h u = u
(E )
As already mentioned, it is easily seen that (u ) is a Palais-Smale sequence for

(E ). If (3.1.1) holds we can therefore apply Theorem 3.1 to (u ). The purpose of this section is to illustrate the idea that, because of (E ), we easily get
complementary information by applying, for instance, the De GiorgiNashMoser
iterative scheme to the u s. We also propose a very simple construction of an
elementary blow-up. First, for the sake of completeness, we recall the content of
the De GiorgiNashMoser iterative scheme. Following standard arguments, as in
Han and Lin [41], we can prove the following.
L EMMA 3.4 Let B0 (3), > 0, be the ball in Rn of center 0 and radius 3, and
let g be a Riemannian metric on B0 (3). Let A > 0 be such that for any smooth
function with compact support in B0 (3),
L2 (B0 (2)) AL2 (B0 (2))
where the norms are taken with respect to g, and u C 1 (B0 (2)), u > 0, be such
that
where
g u f u

B0 (2)
|f | dvg K for some r > n/2. Then, for all p > 0,

r
sup u CuLp (B0 (2))
B0 ()
where the above norm is taken with respect to g, and C = C(n, A, K, p, r, )

depends only on n, A, K, p, r, and .
Another useful version (we refer, for instance, to Han and Lin [41]) is the fol
lowing. With respect to the notations in Han and Lin [41], we let c = h u2 2
and f = 0. We also apply a Trudinger-type argument [74] in order to prove first
that u Lk (M ) for some k > 2 .
L EMMA 3.5 Let (M, g) be a smooth compact Riemannian n-manifold, n 3,
h be a smooth function on M , and u H12 (M ), u 0, be such that for any
nonnegative H12 (M ),

u, dvg +
hudvg
u2 1 dvg .
M
Then u L (M ). Moreover, for any x in M , any > 0, any p > 0, any q > 2 ,
and any > 0, if u is some nonnegative function of H12 (M ) satisfying the above
equation and
uq dvg ,
Bx (2)
46
CHAPTER 3
then
sup u(y) C
yBx ()
Bx (2)
up dvg
1/p
where C > 0 does not depend on u.

As a remark, one passes from small s (given by Han-Lin [41]) to arbitrary s
by writing that

Bx ()
By (y )
where y < . Hence Bx ()
equations
yBx ()
i=1,...,N
Byi (yi ) for some yi Bx (). The
sup u Ci uLp (Bx (2yi ))
Byi (yi )
then give that

sup u CuLp (Bx (2))
Bx ()
since Byi (2yi ) Bx (2).

Unless otherwise stated, we do not assume here that (3.1.1) holds, but assume
that the operators L
g = g + h are uniformly coercive in the sense of (2.1.3). In
other words, we assume that
there exists h smooth, s.t. g + h is coercive, and h h for all .
(3.3.1)
For the sake of simplicity, we also assume that

u 0 weakly but not strongly in H12 (M )
as +. Clearly, u 2 0 as +. We let =
u
= u 1
2 u . Then
2 1
+ h u
= u
g u
4/(n2)
u 2
(3.3.2)
and set
)
(E
and
u 2 = 1. Moreover, the s are bounded, and it easily follows from
(3.3.1) and (3.3.2) that there exists c > 0, independent of , such that for any ,
u ) is bounded in H12 (M ), and that (
u ) also
c. It follows that the sequence (
satisfies (3.3.2). Following standard terminology, we say that x is a (geometrical)
blow-up point for (
u ), or for (u ), if for any > 0,

u
2 dvg > 0
lim sup
+
Bx ()
where Bx () is the geodesic ball of center x and radius . We let S be the set
consisting of the geometrical blow-up points for (
u ). Clearly, S = . We claim
that, up to a subsequence, the following holds:
S is a finite set and
0
u 0 in Cloc
(M \S) .
(3.3.3)
47
We prove (3.3.3). Let x be some point in M , > 0, and 0 1 a smooth

) by
function such that = 1 in Bx (/2) and = 0 in M \Bx (). Multiplying (E
2

, where 1 2 1, we get that
u
2 +1
g u
+ h 2 u
+1
= 2 u
.
2 u
(3.3.4)
Integrating by parts,
+1
4
2
2
u
g u
dvg =
|(
u
)|2 dvg
( + 1)2 M
M
2( 1)
2
+1
(
)
u
dv
||2 u
+1
g
g
dvg
( + 1)2 M
+1 M
while by Holders inequalities,
2 u
2 +1 dvg
+1
2
(
u )2 dvg
22
Bx ()
u
2 dvg
222

.
Moreover, one gets by the Sobolev inequalities that there exist A, B > 0, independent of and , such that
22

+1
+1
2
2
2
2
(
u ) dvg
A
|(
u )| dvg + B
2 u
+1
dvg . (3.3.5)
M
From (3.3.1), there also exists C > 0, independent of and , such that

+1
4
1
|(
u 2 )|2 dvg
C+
( + 1)2
M
+1
4
2
2
|(
u )| dvg +
h 2 u
+1
dvg
( + 1)2 M
M
so that, for any 1 sufficiently close to 1,
+1
+1
C
4
2
2
|(
u
|(
u 2 )|2 dvg
)|
dv
+
h 2 u
+1
g
dvg .
2 M
( + 1)2 M
M
Integrating (3.3.4) then gives that
+1
|(
u 2 )|2 dvg C1 ()
M
+1
|(
u 2 )|2 dvg + C2 (, ) ,
u 2 = 1,
where, since for some > 0, and
2A
C1 () =
C
and
Bx ()
u
2 dvg
222
4( 1)
(g )
u+1
dvg
( + 1)2 C M
4
2B
+
||2 u
+1
dv
+
2 u
+1
g
dvg .
( + 1)C M
C M
C2 (, ) =
(3.3.6)
48
CHAPTER 3
Clearly, there exists at least one geometrical blow-up point for the sequence (
u ).
Let x0 be such a point. We claim that for all > 0,

222

C
2
.
(3.3.7)
u
dvg
lim sup
2A
+ Bx ()
0
If this is not the case, then there exists C1 (0, 1) such that C1 () C1 for all

. By (3.3.6), and since (
u ) is bounded in L2 (M ), it follows that there exists
C2 > 0 such that for any ,
+1
|(
u 2 )|2 dvg C2 .
(3.3.8)
M
By Holders inequalities,

u
2 dvg
Bx0 (/2)
2 (+1)
2
Bx0 (/2)
n+2
n(+1)

dvg
n(+1)
n2
Bx0 (/2)
n2
n(+1)
dvg
so that, with (3.3.8) and according to the Sobolev inequality (3.3.5),

n2

n(+1)
u
2 dvg C3
n(+1)
Bx0 (/2)
Bx0 (/2)
u
n2 dvg
where C3 > 0 is independent of . Since x0 is assumed to be a blow-up point

for the sequence (
u ), we then get the existence of C4 > 0 such that for some
subsequence
n(+1)
(3.3.9)
u
n2 dvg C4 .
Bx0 (/2)
Fix > 1 close to 1, and let 1 = n( + 1)/(n 2). Then 1 < 1 < 2 . Since
u 1 0 as +. Then, by

u 2 0 as +, we also have that
u 1 0 as +.
Holders inequalities, and since
u 2 is bounded,
This contradicts (3.3.9), so that (3.3.7) is proved. Then, as an immediate consequence of (3.3.7) and the relation
u 2 = 1, we get that, up to a subsequence,
(
u ) has a finite number of geometrical blow-up points. Let S be the set of blow-up
points of this subsequence,
and let x now be a point in M \S. There exists > 0

such that u
0 in L2 Bx () . Coming back to (3.3.6), there exists > 1 such
(+1)/2
that
u
is bounded in H12 (M ). Thanks
to the Sobolev embedding theorem,
it follows that u
is bounded in Lq Bx (/2) for some q > 2 . Noting that
2 1
g u
+ h
u u
where h is as in (3.3.1), and applying the De GiorgiNashMoser iterative

scheme
0 as + in C 0 Bx (/4) .
to the u
s (see Lemma 3.5) we get that u
This proves (3.3.3).
As a final remark, let us assume that we also have (3.1.1), in addition to (3.3.1)
and (3.3.2). Then Theorem 3.1 applies to the u s with, in this context, u0 0.
49
xl , l m, be the limits as + of the xj s of Theorem 3.1,

We let x
1 ,. . . ,
j = 1, . . . , m. Let also S be the set consisting of the x
i s, i = 1, . . . , l. We claim
that S = S, where S is as above. As is easily checked, it follows from the proof of
Lemma 3.3 that, for any j = 1, . . . , m,

u2 dvg > 0 .
lim inf
+
j
j (1/R )
x
Hence, for any > 0, and any i = 1, . . . , l,

u2 dvg > 0 .
lim inf
+
Bx
i ()
there exists
This proves that S S. Conversely,
(3.1.3)
gives that for any x S,

2
> 0 such that u 0 in L Bx () . Hence, x S, and S S. This proves

the above claim.
Chapter Four
Exhaustion and Weak Pointwise Estimates
g u + h u = u2
(E )
where u is required to be positive. We let (u ) be a sequence of solutions to (E ),

so that u > 0 and
2 1
g u + h u = u
for all . We describe in this chapter an exhaustion method for the blow-up behavior of the u s. This method provides a constructive approach to Theorem 3.1, and
weak (with respect to the material in Chapters 5 and 6) pointwise estimates on the
u s.
We assume in what follows that there exist 0 < < 1 and a smooth (or only
C 0, ) function h on M such that
the operator g + h is coercive and
(4.0.1)
h h in C 0, (M ) as + .
We also assume that there exists > 0 such that E(u ) for all , where
E(u) = u2 is as in Chapter 2.
Assuming (4.0.1) and that E(u ) for all , it is easily seen that the u s are
bounded in H12 (M ). Indeed, we clearly have that h n/2 C for all , where
C > 0 is independent of . Multiplying (E ) by u and integrating over M , we
then get that

u2 dvg 2 .
|u |2 dvg +
h u2 dvg =
M
From Holders inequality,

2

h u dvg

M
2/n
|h |
n/2

u2 dvg
dvg
2/2
.
It follows that there exists C > 0 such that u 2 C for all , and then that the

u s are bounded in H12 (M ) since they are also bounded in L2 (M ) and 2 < 2 .
Up to a subsequence, we may therefore assume that for some u0 H12 (M ),
u u0 weakly in H12 (M )
0
as +. It is easily seen that u is a solution of

g u0 + h u0 = (u0 )2
(4.0.2)
52
CHAPTER 4
Another easy remark is that (4.0.1) implies the uniform coercivity condition (3.3.1)
for sufficiently large. Indeed, if g +h is coercive, then there exists > 0 such
that g +(h ) is still coercive. Since h h in C 0, (M ), h h for
sufficiently large. In particular, there exists h such that h h for sufficiently
large, and such that g + h is coercive.
Without loss of generality, we can assume in the following that
lim max u = +
+ M
(4.0.3)
so that blow-up occurs. If this is not the case, up to a subsequence, we easily get
from standard regularity theory that u u0 strongly in C 2 (M ) as +.
The main result of this chapter, Theorem 4.1, is stated in section 4.1. Section 4.2
is concerned with the proof of this result.
4.1 WEAK POINTWISE ESTIMATES
We set up some notation. Given N N , let (xi, ), i = 1, . . . , N , be N converging
sequences of points in M , and (i, ), i = 1, . . . , N , be N sequences of positive
real numbers converging to 0. We set

(4.1.1)
S=
lim xi, , i = 1, . . . , N
+
and when N 2 we set

1
lim
exp1
(x
)
,
x
B
()
,
j
=

i
Si =
j,
j,
xi,
xi,
+ i,
(4.1.2)
for i = 1, . . . , N , where 0 < ig /2, ig is the injectivity radius of (M, g), and
the limits, up to a subsequence, are assumed to exist. As in Chapter 3, we regard
expx as defined in Rn . An intrinsic definition is possible if M is parallelizable.
j , j = 1, . . . , k, be open subsets of M such that for any
If not we let j and
j , and such that M = j . The canonical
j, j is parallelizable and j
exponential map gives k maps expx defined in j Rn , and expx is, depending
on the situation, one of these maps. A property of expx that holds for any x M
should then be regarded as a property that
holds
for any j and any x j . We also
let ui, be the function defined in B0 1
i, , the Euclidean ball of center 0 and
1
radius i, , by

n
2 1
ui, (x) = i,
u expxi, (i, x) .
(4.1.3)
Conversely, if u is a smooth function defined in Rn , we let vi, = vi, (u) be the
function given by

1
1 n
1
2
expxi, (x) if x Bxi, () and
vi, (x) = i, u
i,
(4.1.4)
vi, (x) = 0 otherwise .
EXHAUSTION AND WEAK POINTWISE ESTIMATES
53
Finally, we let u = u10 , where u10 is as in Chapter 3, be the function defined in Rn

by
1 n2

|x|2
.
(4.1.5)
u(x) = 1 +
n(n 2)
Then, u is a positive solution, the only one satisfying u(0) = 1 = maxRn u, of the
equation

u = u2 1
where is the Euclidean Laplacian. Its energy E(u) = u2 is E(u) = min ,
(n2)/2
where min = Kn
is as in Chapter 2.
In what follows, we let be a smooth cutoff function in Rn such that 1 in
B0 (/2) and 0 in Rn \B0 (). Then we define i, by i, (x) = (i, x) for
i = 1, . . . , N . When N 2, we consider the following statement:
dg (xi, , xj, )
+
(4.1.6)
min (i, , j, )
N
for all i = j, as +. We let also R
be the function given by
N
R (x) = min dg (xi, , x) .
i=1,...,N
(4.1.7)
The main result of this chapter is the following theorem.

T HEOREM 4.1 Let (M, g) be a smooth compact Riemannian manifold of dimension n 3, (h ) be a sequence of smooth functions on M such that (4.0.1) is satisfied and (u ) be a sequence of positive solutions to (E ) such that E(u )
for some > 0 and all . We assume that (4.0.2) and (4.0.3) hold. Then there
exist N N , converging sequences (xi, ) in M , and sequences (i, ) of positive
real numbers converging to 0, i = 1, . . . , N , such that (4.1.6) holds, and such that,
up to a subsequence, the following propositions hold:
0
(M \S) as +, where u0 is given by (4.0.2),
(P1) u u0 strongly in Cloc
and there exist 0 > 0 and xi Rn such that for any i = 1, . . . , N ,
lim i, (x)ui, (x) = u(x xi )
+
2
weakly in D12 (Rn ), strongly in Cloc
(Rn \Si ), and strongly in C 2 (B0 (0 )), where S
is as in (4.1.1), Si is as in (4.1.2), ui, is as in (4.1.3), and u is as in (4.1.5).
n
N
(x) 2 1 u (x) C, where C > 0 is
(P2) For any x M , and any , R
independent of . Moreover,

n
N
lim
sup
R
(x) 2 1 u (x) u0 (x) = 0
lim
R+ + xM \ (R)
N
where u0 is given by (4.0.2), R
is given by (4.1.7), and (R) is given by
N
(R) = i=1 Bxi, (Ri, ).
(P3) The energy E, given by E(u) = u2 , satisfie
N
2

lim E u u0
vi,
lim E(u )2 u0 22 N 2min
+
i=1
where u0 is given by (4.0.2), vi, = vi, (ui ) is as in (4.1.4), and ui (x) = u(xxi )
where u is as in (4.1.5) and xi is as in (P1).
54
CHAPTER 4
As a remark, because of (P3), we can saturate Theorem 4.1 in order to get some
maximal N by adding sequences (xj, ) and (j, ) such that (P1)(P3) continue to
hold. On the other hand, we clearly have that (4.0.1) implies (3.1.1). We let S be
the set of geometrical blow-up points given by (4.1.1) that we get from Theorem
4.1. We also let S be the set consisting of the limits of the (xik )s we get from
Theorem 3.1. An easy remark is that S and S coincide.
4.2 EXHAUSTION OF BLOW-UP POINTS
We prove Theorem 4.1 in this section. We proceed in several steps. The claims in
what follows are up to a subsequence. First we claim that the following holds.
S TEP 1. There exist a converging sequence (x1, ) of points in M and a sequence
(1, ) of positive real numbers converging to 0 such that, up to a subsequence, the
following propositions hold :
2
(A1) u1, u in Cloc
(Rn ) as +,
(A2) lim u u0 v1, 22 lim E(u )2 u0 22 2min ,

where u1, is as in (4.1.3), u is given by (4.1.5), and v1, = v1, (u) is as in

(4.1.4).
Proof of step 1. We let x1, be a point in M where u achieves its maximum and
set
1 n
u (x1, ) = 1, 2 = max u .
M
From (4.0.3),
lim 1, = 0 .

and for x B0 1
1, , the Euclidean ball of center 0 and
+
We let 0 < < 2g

radius 1
1, , we set

n
2 1
u expx1, (1, x) ,
u1, (x) = 1,
g1, (x) = expx1, g (1, x) ,

h1, (x) = h expx1, (1, x) .
Since 1, 0 as +, it is clear that

2
lim g1, = in Cloc
(Rn )
where is the Euclidean metric. Note that we also have that g1, is controlled on
both sides by in the sense of bilinear forms. It is easily checked that
u1, (0) = u1, L (B0 (1 )) = 1
(4.2.1)
2 1
g1, u1, + h1, 21, u1, = u1,
(4.2.2)
1,
and that
55

1
in B0 1, . We claim now that

2
(Rn )
lim u1, = u1 in Cloc
(4.2.3)
where u1 D12 (Rn ), u1 0, satisfies

u1 = u12
(4.2.4)
where is the Euclidean Laplacian. We also claim that 1, u1, u1 in D12 (Rn )
as +, where 1, is as in the introduction of this section. In order to prove
these claims, we first note that the 1, u1, s are bounded in D12 (Rn ). Indeed, it
is clear that there exists C > 0, independent of , such that |1, | C1, .
Letting s = 2 /(2 2), so that 2s = n, we then write that
|(1, u1, )|2 dvg1,

Rn
2
2
|1, |2 u21, dvg1, + 2
1,
|u1, |2 dvg1,
Rn
Rn
Rn
|1, |2s dvg1,
1/s
2/2
u21, dvg1,
B0 (/1, )
+ O(1)
O(1) .
This proves that the 1, u1, s are bounded in D12 (Rn ). We may therefore assume
that 1, u1, u1 in D12 (Rn ) as + for some u1 D12 (Rn ). From (4.0.1),
0
the coefficients in equation (4.2.2) are bounded and h1, 21, 0 in Cloc
(Rn ) as
+. By standard elliptic theory, using (4.2.1), this gives (4.2.3) and (4.2.4).
It is clear from (4.2.1) and (4.2.3) that u1 (0) = u1 = 1. From CaffarelliGidas-Spruck [17], we necessarily have that u1 = u where u is as in (4.1.5). In
particular, (A1) holds. Let us now prove that (A2) holds also. We write that

0
2
|u u0 v1, |2 dvg
|u u v1, | dvg =
M \Bx1, (R1, )
|u u0 v1, |2 dvg
Bx1, (R1, )
where R > 0 is arbitrary. From (A1),

lim
|u u0 v1, |2 dvg = 0 .
+
Bx1, (R1, )
We write now that for any > 0 there exists C > 0 such that

|u u0 v1, |2 dvg
M \Bx1, (R1, )
(1 + )
+C
M \Bx1, (R1, )
M \Bx1, (R1, )
|u u0 |2 dvg
2
v1,
dvg .
56
CHAPTER 4
Here, C does not depend on R and , and C is defined by the property that for
any x, y R,
|x + y|2 (1 + ) |x|2 + C |y|2 .
Direct computations give that

lim
lim
R+ +
2
v1,
dvg = 0 .
M \Bx1, (R1, )
Combining these relations,

|u u0 v1, |2 dvg
lim
+ M

lim
|u u0 |2 dvg
+ M

lim
lim
|u u0 |2 dvg .
R+ +
From (A1),
lim
lim
R+ +
Bx1, (R1, )
|u u0 |2 dvg = lim
Bx1, (R1, )
R+
u2 dx
B0 (R)
= 2min .
Moreover, since u u0 weakly in H12 (M ), and u u0 almost everywhere,

|u u0 |2 dvg = lim u 22 u0 22 .
lim
+
Thus (A2) is proved thanks to these relations. This ends the proof of step 1.
Let k N, k 1, and for i = 1, . . . , k, let (xi, ) be k converging sequences

of points in M and (i, ) be k sequences of positive real numbers converging to 0.
We consider the following assertions:
(B1) when k 2, (xi, ) and (i, ) satisfy that
+ for any i = j,
dg (xi, ,xj, )
i,
+ as
(B2) for any i {1, . . . , k}, i, ui, u weakly in D12 (Rn ) and strongly in
as +,
2
(Rn )
Cloc
(B3) lim u u0
+
k

i=1
vi, 22 lim E(u )2 u0 22 k2min

where ui, is as in (4.1.3), u is as in (4.1.5), and vi, = vi, (u) is as in (4.1.4).

We let (Hk1 ) be this sequence of assertions and say that (Hk1 ) holds if there exist
k converging sequences (xi, ) of points in M and k sequences (i, ) of positive
real numbers converging to 0 such that, up to a subsequence, (B1)(B3) hold. By
step 1, we know that (H11 ) holds. A second step in the proof of the theorem is the
following:
57

1
) holds or
S TEP 2. Assume that (Hk1 ) holds. Then either (Hk+1
k
R
(x) 2 1 u (x) C
n
k
is given by (4.1.7) and C > 0 is independent
for all x M and all , where R
of x and .
Proof of step 2. Let us assume that

k
(4.2.5)
(x) 2 1 u (x) +
max R
xM
1
holds. We let y M be such that
as +. We need to prove that Hk+1
n
k
k
max R
(x) 2 1 u (x) = R
(y ) 2 1 u (y )
n
xM
(4.2.6)
and we set
1 n
2
.
xk+1, = y and u (xk+1, ) = k+1,
Since M is compact, (4.2.5) implies that k+1, 0 as +. It also follows

from (B2) of (Hk1 ) and (4.2.5) that for any i {1, . . . , k},
dg (xi, , xk+1, )
+
i,

i
as +. Let 0 < < 2g . For x B0 1
k+1, , the Euclidean ball of center
0 and radius 1
k+1, , we set

n
2 1
u expxk+1, (k+1, x) ,
uk+1, (x) = k+1,
gk+1, (x) = expxk+1, g (k+1, x) ,

hk+1, (x) = h expxk+1, (k+1, x) .
Since k+1, 0 as +,
lim gk+1, =
2
in Cloc
(Rn ), where is the Euclidean metric. We also have that gk+1, is on both
sides controled by in the sense of bilinear forms. It is easily checked that
uk+1, (0) = 1
and that
2 1
gk+1, uk+1, + hk+1, 2k+1, uk+1, = uk+1,

n
in B0 1
k+1, . Let x R . For i {1, . . . , k},
dg (xk+1, , xi, ) k+1, |x|

dg xi, , expxk+1, (k+1, x) dg (xk+1, , xi, ) + k+1, |x| .
By (4.2.5), for any i {1, . . . , k},
dg (xi, , xk+1, )
+
k+1,
58
CHAPTER 4
as +, so that
lim

dg xi, , expxk+1, (k+1, x)
dg (xi, , xk+1, )
=1
for all i {1, . . . , k}. From (4.2.6), it follows that

lim sup uk+1, (x) 1
+
for all x R . As in step 1, it is easily seen that, up to a subsequence,

n
2
(Rn )
lim uk+1, = uk+1 in Cloc
where uk+1 0 is such that

2 1
uk+1 = uk+1
and where is the Euclidean Laplacian. We also get that k+1, uk+1, uk+1
weakly in D12 (Rn ) as +. From Caffarelli-Gidas-Spruck
[17], uk+1 = u
1
. Independently, we have
where u is as in (4.1.5). This proves (B2) of Hk+1
1
already seen that (B1) of Hk+1
is true. Thus, it remains to prove that (B3) of
1
Hk+1 is also true. Let R > 0 be given. From the expression for u, there exists
C > 0, independent of , such that for any i {1, . . . , k},
2n
2
vi,
dvg Cni,
dg (xi, , x)
dvg .
Bxk+1, (Rk+1, )
Bxk+1, (Rk+1, )
1
Using (B1) of Hk+1
, it is easily checked that for large enough and for any
x Bxk+1, (Rk+1, ),
2n
dg (xi, , x)
Thus,
2n
2dg (xi, , xk+1, )
2
vi,
dvg
Bxk+1, (Rk+1, )

2n n
2Cdg (xi, , xk+1, )
i, Volg Bxk+1, (Rk+1, )
nk+1,
ni,
C(R)
n
n
dg (xi, , xk+1, ) dg (xi, , xk+1, )
where C(R) is independent of , and for M , Volg () stands for the volume
1
, it follows that
of with respect to g. Using again (B1) of the sequence Hk+1
for any i {1, . . . , k} and any R > 0,
2
lim
vi,
dvg = 0 .
(4.2.7)
+
Bxk+1, (Rk+1, )
1
From (B2) of Hk+1
, we also have that
lim
lim
R+ +
Bxk+1, (Rk+1, )
u2 dvg = 2min .
(4.2.8)
59
We write now that

2
k

0
vi, dvg
u u

M \Bxk+1, (Rk+1, )
i=1

k
2

0
=
vi, dvg
u u

M
i=1
2

k

0
vi, dvg .
u u

Bx
(Rk+1, )
i=1
k+1,
Since

2
k

0
vi, dvg
u u

Bxk+1, (Rk+1, )
i=1
u2 dvg + 2
Bxk+1, (Rk+1, )
+2
Bxk+1, (Rk+1, )

i=1
2 1 0
u
u dvg
2 1
u
vi, dvg ,
Bxk+1, (Rk+1, )

it follows from (B3) of Hk1 , (4.2.7), and (4.2.8) that
2

k

lim
vi, dvg
lim
u u0
R+ + M \B

(Rk+1, )
x
(4.2.9)
i=1
k+1,
2

u0 22

1)2min .
(k +
lim E(u )
+
1
Using (B2) of Hk+1
, we also have that

|u vk+1, |2 dvg = 0
lim
(4.2.10)
and direct computations give that
lim
lim
(4.2.11)
Bxk+1, (Rk+1, )
R+ +
Writing that
Bxk+1,
M \Bxk+1, (Rk+1, )
i=1
|u vk+1, |2 dvg
Bxk+1, (Rk+1, )
k

i=1
21
2

k+1

vi, dvg
u u 0

(Rk+1, )
2
vk+1,
dvg = 0 .
Bxk+1, (Rk+1, )
2
vi,
dvg
21

21
+
Bxk+1, (Rk+1, )
0 2
(u ) dvg
21
,
60
CHAPTER 4
we get with (4.2.7) and (4.2.10) that

2
k+1

lim
vi, dvg = 0 .
u u0
+ B

(R
)
x
k+1,
(4.2.12)
i=1
k+1,
Given > 0 and R > 0, we write that

2

k+1

0
vi, dvg
u u

M
i=1
2

k+1

vi, dvg
u u0

Bxk+1, (Rk+1, )
i=1

2
k

0
+ (1 + )
vi, dvg
u u

M \Bxk+1, (Rk+1, )
i=1
2
+C
vk+1,
dvg
M \Bxk+1, (Rk+1, )
where C does not depend on R and . Passing to the limit in this expression as
+, then as R + and finally as 0, we get from (4.2.9), (4.2.11),
and (4.2.12), that
2

k+1

vi, dvg lim E(u )2 u0 22 (k+1)2min .
lim
u u0
+ M
+

i=1
1
Hence, (B3) of Hk+1
is true, and this ends the proof of Step 2.
2
Following steps 1 and 2, by induction, we get that there exists k 1, with

2min k 2 u0 22 , that there exist k converging sequences (xi, ) of points in
M and that there exist k sequences (i, ) of positive real numbers converging to
0, such that (B1)(B3) hold, and such that for any x M and any ,
k
R
(x) 2 1 u (x) C
n
(4.2.13)
where C > 0 is independent of and x. By (4.2.13), u L () C for any

M \S, where S, given by (4.1.1), consists of the limits of the xi, s, and
C depends only on . An easy claim then is that
0
lim u = u0 in Cloc
(M \S) .
This easily follows from standard elliptic theory.

Let k N, k 1, and for i {1, . . . , k}, let (xi, ) be k converging sequences
of points in M , and (i, ) be k sequences of positive real numbers converging to
0. We consider the following assertions:
(C1) when k 2, (xi, ) and (i, ) satisfy that
i = j,
dg (xi, ,xj, )
min{i, ,j, }
0
(C2) lim u = u0 in Cloc
(M \S) where S is as in (4.1.1),
+
+ for any
61
(C3) there exists 0 > 0 and there exists xi Rn , i = 1, . . . , k, such that for any
2
(Rn \Si ),
i = 1, . . . , k, i, ui, u ( xi ) weakly in D12 (Rn ), strongly in Cloc
2
and strongly in C (B0 (0 )) as +, where ui, is as in (4.1.3), Si is as in
(4.1.2), and u is given by (4.1.5),
k
(C4) for any x M and any , R
(x) 2 1 u (x) C where C > 0 is independent of x and ,
k

(C5) lim u u0 i=1 vi, 22 lim E(u )2 u0 22 k2min
n
where vi, = vi, (ui ) is as in (4.1.4), and ui (x) = u(x xi ), where u is as in

(4.1.5) and xi is as in (C4).
We let (Hk2 ) be this sequence of assertions and say that (Hk2 ) holds if there
exist k converging sequences (xi, ) of points in M and k sequences (i, ) of
positive real numbers converging to 0 such that, up to a subsequence, (C1)(C5)
hold. According to what we just said, as a consequence of steps 1 and 2, there
exists k 1 such that (Hk2 ) holds. A third step in the proof of the theorem is the
following:
2
) holds or
S TEP 3. Assume that (Hk2 ) holds. Then either (Hk+1

n
k
1

lim
lim
sup
R (x) 2
u (x) u0 (x) = 0
R+ + xM \ (R)
(4.2.14)
where (R) = ki=1 Bxi, (Ri, ).

2
)
Proof of step 3. We assume that (4.2.14) is false. We need to prove that (Hk+1
holds. Since (4.2.14) is assumed to be false, there exists a sequence (xk+1, ) of
points in M such that for any i = 1, . . . , k,
dg (xi, , xk+1, )
+
i,
(4.2.15)
as +, and such that

k
(xk+1, ) 2 1 |u (xk+1, ) u0 (xk+1, ) | (40 )
R
n
n2
2
(4.2.16)
for some 0 > 0. We let

1 n
2
u (xk+1, ) = k+1,
2
and claim that Hk+1
holds when adding (xk+1, ) and (k+1, ) to the (xi, )s
2
is a consequence
and (i, )s, i = 1, . . . , k. An easy remark is that (C1) of Hk+1
of (4.2.15). We also have that
k
R
(xk+1, ) 0
2
as +. If not, (C2) of Hk implies that
(4.2.17)
|u (xk+1, ) u0 (xk+1, ) | 0
as +, contradicting (4.2.16) since M is compact. Using (4.2.17), we can
rewrite (4.2.16) as
k
(xk+1, )
R
20
k+1,
(4.2.18)
62
CHAPTER 4
for large enough. This clearly implies that

k+1, 0 as + .

We let 0 < < ig /2, and, for x B0 1
k+1, , the Euclidean ball of center 0
and radius 1
k+1, , we set

n
2 1
uk+1, (x) = k+1,
u expxk+1, (k+1, x) ,
gk+1, (x) = expxk+1, g (k+1, x) ,

hk+1, (x) = h expxk+1, (k+1, x) .
Since k+1, 0 as +,
2
lim gk+1, = in Cloc
(Rn ) .
(4.2.19)
We also have that gk+1, is controlled on both sides by in the sense of bilinear
forms. It is easily checked that
uk+1, (0) = 1
and that
2 1
gk+1, uk+1, + hk+1, 2k+1, uk+1, = uk+1,
(4.2.20)

in B0 1
k+1, . We let

1
lim
exp1
(x
)
,
x
B
()
,
1
k
Sk+1 =
i,
i,
xk+1,
xk+1,
+ k+1,
where, up to a subsequence, the limits are assumed to exist. From (4.2.18),

3
Sk+1 B0
0 = .
(4.2.21)
2
Let R > 0 and let (x ) be a sequence of points in B0 (R) such that
d (x , Sk+1 )
1
R
where d is the Euclidean distance. From (4.2.19),

k+1,
k
expxk+1, (k+1, x )
R
2R
for sufficiently large. Letting y = expxk+1, (k+1, x ), it follows from (C4)

of Hk2 that
n
uk+1, (x ) (2R) 2
k
R
(y ) 2
u (y ) (2R) 2
C.
Hence, for any K R \Sk+1 , there exists CK > 0, independent of , such that
n
uk+1, L (K) CK .
(4.2.22)
63
Using (4.2.21) and (4.2.22), we can apply the De GiorgiNashMoser iterative

scheme as stated in Lemma 3.4 to equation (4.2.20). This gives the existence of
some C > 0 independent of , such that

12
1 = uk+1, (0) C
B0 (0 )
u2k+1, dvgk+1,
As in steps 1 and 2, it is easily seen that, up to a subsequence,

2
(Rn \Sk+1 )
lim uk+1, = uk+1 in Cloc
where uk+1 0 is such that

2 1
uk+1 = uk+1
and where is the Euclidean Laplacian. We also get that k+1, uk+1, uk+1
weakly in D12 (Rn ) as +. Thanks to Caffarelli-Gidas-Spruck [17],
(n2)/2
uk+1 (x) = k+1
u (k+1 (x xk+1 ))
where u is as in (4.1.5), k+1 > 0, and xk+1 Rn are such that

k+1 = 1 +
2k+1 |xk+1 |2
.
n(n 2)

2
, up to
Since Si with respect to Hk2 is a subset of Si with respect to Hk+1
2
1
changing k+1, into k+1 k+1, and xk+1 into k+1 xk+1 , (C3) of Hk+1 is
proved. By (4.2.17), the limit as + of xk+1, belongs
to2 S,
where S is the
s,
i
=
1,
.
.
.
,
k.
Hence,
(C2)
of
H
set
of
the
limits
of
the
x
i,
k implies (C2) of
2
Hk+1 . Noting that
min
i=1,...,k+1
dg (xi, , x) min dg (xi, , x) ,
Hk2
i=1,...,k
2
implies (C4) of Hk+1
. As already mentioned,
we also have that (C4) of
2
(C1) of Hk+1 follows from (4.2.15). We are thus left with the proof that (C5) of
2
Hk+1 holds. Let R > 0. For any > 0, there exists C independent of R and
such that

2
k+1

vi, dvg
u u0

M \Bxk+1, (Rk+1, )
i=1

2
k

0
(1 + )
vi, dvg
u u

M \Bxk+1, (Rk+1, )
i=1
2
+C
vk+1,
dvg .
M \Bxk+1, (Rk+1, )
Noting that
lim
lim
R+ +
M \Bxk+1, (Rk+1, )
2
vk+1,
dvg = 0 ,
64
CHAPTER 4
we easily get that

2
k+1

vi, dvg
u u0

M \Bxk+1, (Rk+1, )
i=1

2
k

0
vi, dvg + R ()
u u

M \Bx
(Rk+1, )
(4.2.23)
i=1
k+1,
where
lim
lim R () = 0 .
R+ +
Letting

2
k+1

vi, dvg
IR () =
u u0

Bxk+1, (Rk+1, )
i=1

2
k

0
vi, dvg ,
u u

Bx
(Rk+1, )
(4.2.24)
i=1
k+1,
it follows from (4.2.23) that

2
2

k+1
k

vi, dvg
vi, dvg
u u0
u u0

M
M
i=1
i=1
(4.2.25)
+ IR () + R () .
Let

2
2
k+1
k

F = uk+1, u0
vi, ,
vi, uk+1, u0

i=1
i=1
where

n
2 1
vi, (x) = k+1,
vi, expxk+1, (k+1, x)
and

n
2 1
u0 (x) = k+1,
u0 expxk+1, (k+1, x) .
It is easily checked that
IR () =
Let
B0 (R)

VR =
F dvgk+1, .
1
x R s.t. d (x, Sk+1 )
R
where d is the Euclidean distance. Similar arguments to the ones we used to prove
(4.2.23) give that
F dvgk+1, R () .
B0 (R)VR
Hence
65
IR ()
B0 (R)\VR
F dvgk+1, + R () .
(4.2.26)
Given i {1, . . . , k}, either dg (xi, , xk+1, ) Ck+1, for some C > 0 independent of , or
dg (xi, , xk+1, )
+
k+1,
as +. In the first case, we get by (4.2.15) that
k+1,
+
i,
as +. Hence, given i {1, . . . , k}, either
dg (xi, , xk+1, )
+
k+1,
(4.2.27)
k+1,
+
i,
(4.2.28)
as +, or
as +. If i {1, . . . , k} is such that (4.2.27) holds, we get as in step 2 that
2
vi,
dvg 0
Bxk+1, (Rk+1, )
as +. In particular, for any R > 0,
2
vi,
dvgk+1, 0
B0 (R)\VR
(4.2.29)
as +. Let now i {1, . . . , k} be such that (4.2.28) holds and (4.2.27) does
not hold. Let also x B0 (R) \VR . For sufficiently large,

1
k+1, ,
dg xi, , expxk+1, (k+1, x)
2R
and it follows from the expression of vi, that

n2 1
i,
sup vi, CR
k+1,
B0 (R)\VR
for some CR > 0 independent of . In particular, using (4.2.28), we get that for
any R > 0,
sup
B0 (R)\VR
vi, 0
2
, we also have that for any R > 0,
as +. From (C3) of Hk+1

|uk+1, vk+1, |2 dvgk+1, 0
B0 (R)\VR
(4.2.30)
(4.2.31)
66
CHAPTER 4
as +. Combining (4.2.26) and (4.2.29)(4.2.31), we then get, as in step 2,

that
IR () 2min + R ()
(4.2.32)
where
lim
lim R () = 0 .
R+ +

From (4.2.24), (4.2.25), (4.2.32), and (C5) of Hk2 , it follows that
2

k+1

lim
vi, dvg lim E(u )2 u0 22 (k + 1)2min
u u0
+ M
+

i=1
2
and (C5) of Hk+1 holds. This ends the proof of step 3.
2
We are now in position to prove Theorem 4.1. As already mentioned, it follows
from steps 1 and 2 that there exists k N,
2 u0 22
1k
,

2min
and that there exist k converging sequences (xi, ) of points in M

and k sequences
(i, ) of positive real numbers converging to 0, such that Hk2 holds. From step
3, we then get by induction that there exists N N,
2 u0 22
1kN
,

2min
2
holds and such that
such that HN
lim
lim
sup
R+ + xM \ (R)

n
N
R
(x) 2 1 u (x) u0 (x) = 0
where (R) = N
i=1 Bxi, (Ri, ). This proves Theorem 4.1.
Chapter Five
Asymptotics When the Energy Is of Minimal Type
g u + h u = u2
(E )

so that u > 0 and
2 1
g u + h u = u
for all . For the readers convenience, we describe in this chapter the C 0 -theory
for the blow-up behavior of the u s when the energy of the u s is of minimal
type. The general situation of arbitrary energies is treated in the next chapter.
As in Chapter 4, we assume in what follows that there exist 0 < < 1 and a
smooth (or only C 0, ) function h on M such that
(5.0.1)
h h in C 0, (M ) as + .
We also assume in this section that the energy of the u s is of minimal type in the
sense that
E(u ) min
(5.0.2)
(n2)/2
Kn
for all , where E(u) = u2 and min =

are as in Chapter 2. It
is easily seen that the u s are bounded in H12 (M ). Assuming (5.0.1), we clearly
have that h n/2 C for all , where C > 0 is independent of . Multiplying
(E ) by u and integrating over M , we then get that

2
2
|u | dvg +
h u dvg =
u2 dvg 2min .
M
From Holders inequality,

2

h u dvg

M
2/n
|h |
n/2
dvg

u2 dvg
2/2
.
It follows that there exists C > 0 such that u 2 C for all , and then

that the u s are bounded in H12 (M ) since they are also bounded in L2 (M ) and
we have 2 < 2 . Up to a subsequence, we may therefore assume that for some
u0 H12 (M ), u u0 in H12 (M ) as +.
Assuming (5.0.1) and (5.0.2), one can prove that, up to a subsequence, either
u0 0, and u u0 strongly in C 2 (M ) as +, or u0 0 and u 0
weakly and not strongly in H12 (M ) as +. This is the subject of Proposition
5.1 of section 5.1. Sharp estimates on the u s are then proved in section 5.2.
68
CHAPTER 5
5.1 STRONG CONVERGENCE AND BLOW-UP

We assume (5.0.1) and (5.0.2). As already mentioned, up to a subsequence, we
have that u u0 in H12 (M ) for some u0 H12 (M ) as +. We claim that
the following proposition holds.
P ROPOSITION 5.1 Under the above conditions (5.0.1) and (5.0.2), up to a subsequence, either u0 0, and u u0 strongly in C 2 (M ) as +, or u0 0,
and u 0 weakly and not strongly in H12 (M ) as +.
Proof. An easy remark (we refer to Chapter 4) is that (5.0.1) implies the uniform
coercivity condition (3.3.1) for sufficiently large. In particular, up to a subsequence, there exists h such that h h for all , and such that g + h is coercive.
Then, from the Sobolev embedding theorem,

u2 dvg =
|u |2 dvg +
h u2 dvg
M
M
2
|u | dvg +
hu2 dvg
M
M

C1
|u |2 dvg +
u2 dvg
M
u2 dvg
C2
2/2
where C1 , C2 > 0 are independent of , so that there exists c > 0 such that
u 2 c for all . This proves that, if u0 0, then u 0 weakly and
not strongly in H12 (M ) as +. In the following we assume that u0 0.
A possible proof that the convergence u u0 holds in C 2 (M ) is as follows.
Thanks to regularity theory, it is easily seen that we need to prove only that the
u s are bounded in C 0 (M ). We proceed by contradiction and assume that, up to
4/(n2)
a subsequence, maxM u + as +. We let = u 2
and
1
u
= u 2 u . Then
2 1
g u
+ h u
= u
and
u
2 dvg = 1 .
As already mentioned, there exists c > 0 such that u 2 c. We also have that
Kn2 for all , and still up to a subsequence, we can assume that as
weakly in H12 (M ) as +, where u
is given by
+, and that u
u
(n2)/4 0
u . We let x M and > 0 be such that
u
=
1 n
2
max u
= u
(x ) =
M
Then 0 as +. We use below only the fact that there exists r0 > n/2
such that h
r0 C for all , where C > 0 is independent of , and h is the
69
ASYMPTOTICS WHEN THE ENERGY IS OF MINIMAL TYPE
negative part of h . This easily follows from (5.0.1). Given 0 < < ig , where ig
is the injectivity radius of (M, g), and x Rn such that |x| , we define
n2

expx ( x)
u
(x) = 2 u
where expx is the exponential map at x . Then
2 1
u
g u
+ h
= u

(x) = 2 h exp ( x) and g (x) = (exp ) g( x). For any
where h
x
x
R > 0, and sufficiently large,
2
|
u | dvg
|
u |2 dvg C and
B0 (R)
M

u
2 dvg
u
2 dvg C
B0 (R)
where C > 0 is independent of and R, and B0 (R) is the Euclidean ball of

center 0 and radius R. Let be a smooth cutoff function, 0 1, such that
= 1 in B0 (/2) and = 0 in Rn \B0 (3/4). We let (x) = ( x). Then,
| | C for all . Let s = 2 /(2 2). Then 2s = n, and we can write that
|( u
)|2 dvg 2
| |2 u
2 dvg + 2
2 |
u |2 dvg
Rn
Rn
Rn
2s
1/s
| | dvg
B0 (3/4 )
Rn
2/2

u
2 dvg
+ O(1) O(1) .
s are bounded in D12 (Rn ), where D12 (Rn ) is the homogeIt follows that the u
neous Sobolev space defined as the completion of C0 (Rn ) with respect to the
norm uD12 = u2 . Up to a subsequence, we can therefore assume that
u in D12 (Rn ) as +, u D12 (Rn ), u 0. In particular, u
u
u

in L2 (Rn ) as +, and thus
2 1 .
u2 lim inf u
+
(5.1.1)
We claim now that u 0. We can write that

2 2
| u
g u
.
u
+ |h
r0

Moreover, u
1, and B0 (1) h
dvg C for all , where C > 0 is indepen
dent of , and r0 > n/2 is as above. Hence, thanks to the De GiorgiNashMoser
iterative scheme, as stated in Lemma 3.4,
C
u L2 (B0 (1)) .
1 = max u
B0 (1/2)
It clearly follows that u 0. This proves the above claim. We let C0 (Rn ),
0, and R > 0 such that supp B0 (R). For large,
2 1
dvg (5.1.2)
(
u ) dvg +
dvg
h u
B0 (R)
B0 (R)
B0 (R)
70
CHAPTER 5
and we can write that
B0 (R)
2 rn
|
|h
C
u dvg
h
r0 = o(1)
where C > 0 is independent of , and r0 > n/2 is as above. Hence, passing to the
limit as + in (5.1.2), we get that, for all C0 (Rn ),

1
(u) dx 2
u2 1 dx .
(5.1.3)
K n Rn
Rn
By density, (5.1.3) is true for all D12 (Rn ), 0. Taking = u, it follows
that

1
2
|u| dx 2
u2 dx .
K
n
n
R
n R
Noting that u 0, using the sharp Euclidean Sobolev inequality and (5.1.1), we
can write that

1 22

|u|2 dx
1
1
1
2
Rn
u
dx
2.

2/2
2

Kn2
K
K
n
2
R
n
n
dx
n u
R
2
Hence, Rn u dx = 1, and u is an extremal function for the sharp Euclidean

Sobolev inequality. In particular,
2
u2 lim sup u
+
and, since L2 is uniformly convex, we get that u

u strongly in L2 (Rn ) as
+. Let D M be a measurable set. Noting that there exists C > 0, such
that for any x, y R,

|x + y|2 |x|2 |y|2 C |x|2 1 |y| + |y|2 1 |x|
and that
|
u u
|2
lim
we get that
|
u u
|2 dvg =
|
u|2
|
u|dvg = lim
|
u u
|dvg = 0 ,
u
2 dvg + o(1) .
Taking D = M , it follows that

|
u u
|2 dvg = 1
M
u
2 dvg
u
2 dvg + o(1) .
Taking D = Bx (R ), R > 0, it follows that

2
u
dvg =
|
u u
|2 dvg + o(1) .
Bx (R )
Bx (R )
Now we can write that
B0 (R)
2
u dx
| u
|2 dvg + o(1)
B0 (R)
71
u
2 dvg + o(1)
Bx (R )
|
u u
|2 dvg + o(1)
Bx (R )
u
2 dvg + o(1) .
1
M
Passing to the limit as +, and then as R +, it follows that

11
u
2 dvg ,
M
and we get the contradiction we were looking for. This ends the proof that the
u
s, and thus the u s, are bounded in C 0 (M ). From regularity theory, and up to
a subsequence, we then get that u u0 strongly in C 2 (M ) as +. This
proves Proposition 5.1.
2
From now on we assume that u 0 weakly and not strongly in H12 (M ) as
+. As above, we assume in addition that (5.0.1) and (5.0.2) hold. Since
(5.0.1) implies the uniform coercivity condition (3.3.1) for large, there exists
4/(n2)
c > 0, independent of , such that u 2 c for all . We let = u 2
1
and u
= u 2 u . Then
2 1
+ h u
= u
g u
and
)
(E
u
2 dvg = 1 .
Kn2
for all , while u

0 weakly and not strongly in H12 (M ) as
Clearly,
+. An easy claim is as follows: up to a subsequence, there exists one point
x0 in M such that for any > 0,

lim inf
(5.1.4)
u
2 dvg = 1
+
Bx0 ()
and such that

0
(M \{x0 })
u
0 in Cloc
(5.1.5)
as +. In other words, up to a subsequence, the u

s have one and only one
geometrical blow-up point. In order to prove this claim, we proceed as follows.
As already mentioned, (5.0.1) implies the uniform coercivity condition (3.3.1) for
large. From subsection 3.3 we then get that if S is the set of the geometrical
0 in
blow-up points of (
u ), then, up to a subsequence, S is finite and u
0
(M \S). On the other hand, we clearly have that maxM u
+ as the
Cloc
parameter +. The above argument, used to prove Proposition 5.1, then
72
CHAPTER 5
shows that, up to a subsequence, there exists R > 0, with the property that R 0
as R +, such that for any R > 0,

1 R
u
2 dvg + o(1)
(5.1.6)
Bx (R )
where o(1) 0 as +, and x , are as above. Still up to a subsequence,

we can assume that x x0 as +. Given > 0, and passing to the limit
as + in (5.1.6), we then get that for any R > 0,

u
2 dvg
1 R lim inf
+
Bx0 ()
lim sup
+
u
2 dvg 1 .
Bx0 ()
Passing to the limit as R +, it follows that for any > 0,

lim
u
2 dvg = 1 .
+
Bx0 ()
In particular, S = {x0 }, and (5.1.4)(5.1.5) are proved.

5.2 SHARP POINTWISE ESTIMATES
We let (u ) be a sequence of positive solutions to (E ). We assume that (5.0.1)
and (5.0.2) hold. We also assume (see the discussion in section 5.1) that
u 0 weakly in H12 (M ) as + .
(5.2.1)
We define x M and > 0 by the relations

1 n
2
u (x ) = max u =
M
Clearly, x x0 and 0 as +, where x0 is as in (5.1.4), (5.1.5). We

know from Chapter 4 that there exists C > 0, independent of , such that for any
and any x M ,
dg (x , x)
n2
2
u (x) C
(5.2.2)
and that
lim
lim
sup
R+ + M \Bx (R )
dg (x , x)
n2
2
u (x) = 0 ,
(5.2.3)
where dg is the distance with respect to g. We let u = u10 , where u10 is as in Chapter
3, be the function defined in Rn by
1 n2

|x|2
.
u(x) = 1 +
n(n 2)
Then, u is a positive solution, the only one satisfying u(0) = 1 = maxRn u, of the
equation
u = u2
73
where is the Euclidean Laplacian. Its energy E(u) = u2 is E(u) = min ,
n2
2
(n2)/2
is as in Chapter 2. The standard bubble

where min = Kn
is then given by the expression
n2
2

n2
x
u
.
=
2
2 + |x|
u(1
x)
n(n2)
We claim here that the following sharp estimates hold.

T HEOREM 5.2 Under the above assumptions (5.0.1), (5.0.2), and (5.2.1), up to
a subsequence, there exists C > 1, independent of , such that for any and any
x M,
n2
n2
2
2
u (x) C
d (x ,x)2
C 2 + dg (x ,x)2
2 + g
n(n2)
n(n2)
where x and are as above. Moreover, for any > 0, there exists > 0 such
that, up to a subsequence, for any and any x Bx0 ( ),
n2
n2
2
2
u (x) (1 + )
.
d (x ,x)2
1 + 2 + dg (x ,x)2
2 + g
n(n2)
n(n2)
In particular, the u s are C controlled, on both sides, by the standard bubble.

Proof. We prove this theorem by coming back to the u
s. As in the preceding
4/(n2)
and u
= u 1
u
. Then,
section, we let = u 2
2
2 1
+ h u
= u
g u
and
u
2 dvg = 1
)
(E
(5.2.4)
for all . We know (see the proof of Proposition 5.1) that there exists c > 0,
0 weakly in
independent of , such that u 2 c for all . Clearly, u
H12 (M ) as +, and (5.1.4), (5.1.5) hold. An easy claim is that
1
lim = 2 .
(5.2.5)
+
Kn
On the one hand, we know that Kn2 for all . On the other hand, since the
0 strongly
embedding of H12 (M ) in L2 (M ) is compact, we can assume that u
in L2 (M ). Thanks to the sharp Sobolev inequality, as proved by Hebey-Vaugon
[47, 48], there exists B > 0, independent of , such that

u 22 Kn2
u 22 + B
u 22 .
) and (5.2.4) we can then write that

Using (E

h u
2 dvg Kn2
u 22 + B
u 22 .

u 22 +
M
74
CHAPTER 5
It follows that

u 22
u 22 C
1 Kn2
for all , where C > 0 is independent of . Clearly, there exists c > 0 such that
s converge strongly to

u 2 c. If this is not the case, a subsequence of the u
0 in H12 (M ), contradicting (5.2.4). Passing to the limit as + in the above
equation, this proves (5.2.5).
We define
> 0 by the relations
1 n
2
u
(x ) = max u
M
2/(n2)
where x is as above. Clearly,

= u 2
and
0 as +.
By (5.2.2), (5.2.3), and (5.2.5), we have that that there exists C > 0, independent
of , such that for any and any x M ,
dg (x , x)
n2
2
u
(x) C
and that
lim
lim
sup
R+ + M \Bx (R
)
dg (x , x)
n2
2
u
(x) = 0
(5.2.6)
where dg is the distance with respect to g. It is easily checked that the equations in
Theorem 5.2 reduce to the existence of C > 1, independent of , such that for any
and any x,
n2
n2
2
2
1
u
(x) C
2/n
2/n
d (x ,x)2
d (x ,x)2
C
2 + n g
2 + n g
(5.2.7)
and, for any > 0, to the existence of > 0, independent of , such that for any
and any x Bx0 ( ),
n2
n2
2
2
(x)
2/n
2/n
d (x ,x)2
d (x ,x)2
C
2 + n g4
2 + n g4
(5.2.8)
where C = 1 + . We prove (5.2.7) and (5.2.8) in what follows. The proof
proceeds in several steps. We claim first that there exists C > 0 such that, up to a
subsequence,

n2
2
u
(x) C
2/n
2 + n4 dg (x , x)2
for all and all x M . It is easily checked that the proof of this claim reduces to
the proof that there exists C > 0 such that for any and any x M ,
(x )
u (x) C .
dg (x , x)n2 u
(5.2.9)
75
As a first step we prove that for any > 0 there exists C > 0 such that for any
and any x M ,
n2
2 +
dg (x , x)n2
u
(x) C .
(5.2.10)
In step 2 we prove that we can take = 0 in (5.2.10), so that (5.2.9) is true. Step
3 is concerned with the exact asymptotic profile of the u
s. The proof of (5.2.7)
and (5.2.8), and thus of Theorem 5.2, follows from this exact asymptotic profile.
S TEP 1. We prove that (5.2.10) is true. An easy claim is that it suffices to prove
(5.2.10) for > 0 small. Indeed, let 1 < 2 be positive. It is easily seen that
. On the other hand, if
(5.2.10) with respect to = 1 , 2 is true if dg (x , x)
, then
(5.2.10) is true with respect to 1 , and dg (x , x)
n2 +
2
2
u
(x)
dg (x , x)n22

n2 +1
2
1
dg (x , x)1
dg (x , x)n21
2
u
(x)
so that (5.2.10) with respect to 2 is also true. This proves the above claim. As
already mentioned, (5.0.1) implies the uniform coercivity condition (3.3.1). We
let h C 0, (M ) be such that, up to a subsequence, h h for all , and g + h
0
is coercive. Clearly, there exists 0 > 0 such that for > 0 small, g + h
1 is
coercive. We fix > 0 small, and let G be the Greens function of this operator.
By the maximum principle, G is positive. We let L be the operator
2 2
u.
L u = g u + h u u
= 0 with u
> 0 on M , we get from Berestycki-Nirenberg-Varadhan
Since L u
[9] that the maximum principle holds for L . It is easily seen that in M \{x },
(x , x)
L G1
|G |2 (x , x)
2 2
=
+
h
(x)
h(x)
(x)
+
(1
)
0
G2 (x , x)
G1
(x , x)
so that
(x , x)
L G1
|G |2 (x , x)
2 2
(x)
+
(1
)
0
G2 (x , x)
G1
(x , x)
(5.2.11)
A standard property of the Greens function we use in the following (we refer to
Appendix A for more details) is that there exist C > 0 and > 0 such that, for any
and any x Bx ()\{x },
|G |(x , x)
C
.
G (x , x)
dg (x , x)
(5.2.12)
Let R > 0, to be fixed later on. Combining (5.1.5) and (5.2.11), we get that for
sufficiently large
(x , x) 0
L G1
), then, from (5.2.6),

in M \Bx (). On the other hand, if x Bx ()\Bx (R
2 2
(x) R
dg (x , x)2 u
76
CHAPTER 5
where R 0 as R +. Together with (5.2.11), (5.2.12), this gives that

(x , x)
L G1
C(1 ) R
0 +
.
1
dg (x , x)2
G (x , x)
We choose R > 0 sufficiently large such that C(1 ) R 0. Then
(x , x) 0 in Bx ()\Bx (R
), and we have proved that for suffiL G1
(x
,
x)
0
in
M
\B
). By another standard property
ciently large, L G1
x (R
of the Greens function (we refer here again to Appendix A for more details) there
),
exists C > 0 such that for any , and any x Bx (R
(x , x) C
(1)(n2)
.
G1
(1)(n2) n2
2
If we let C = C 1
, we then get that
(x , x) u
(x)
C G1
). From the maximum principle, this implies that

for all and all x Bx (R
(x , x) u
(x)
C G1
). Noting that there exists C > 0 such that for

for all , and all x M \Bx (R
any , and any x M \{x },
dg (x , x)n2 G (x , x) C ,
it follows that for any > 0, any , and any x M \{x },
n2
2 +
dg (x , x)n2
u
(x) C
where = (n 2), and C > 0 is independent of . In particular, (5.2.10) is true

). Noting that (5.2.10) is obviously satisfied in Bx (R
), this
in M \Bx (R
proves (5.2.10).
S TEP 2. We prove that we can take = 0 in (5.2.10) so that (5.2.9) is true.
(x )
u (x) is maximum. We need to
We let y be a point where dg (x , x)n2 u
(x )
u (y ) is bounded as +. We let G be the
prove that dg (x , y )n2 u
Greens function of the operator g + h, where h is as above. Since G is positive,
G(y , x) (g u
+ h
u ) (x)dvg (x)
u
(y ) =
G(y , x) (g u
+ h u
) (x)dvg (x)
M
2 1
=
G(y , x)
u
(x)dvg (x) .
M
Given > 0 small, we write that
2 1
G(y , x)
u
(x)dvg (x)
M
2 1
=
G(y , x)
u (x)dvg (x) +
Bx ()
M \Bx ()
2 1
G(y , x)
u
(x)dvg (x) .
77
Using (5.2.10), we can write that
2 1
G(y , x)
u
(x)dvg (x)
M \Bx ()
( n2 )(2 1)
=O

n+2

2 (2 1)
M \Bx ()
G(y , x)dvg (x)

G(y , x)dvg (x)

=O
M
n+2

(2 1)
=O
2
.
Therefore,
u
(y ) C
Bx ()

n+2
(2 1)
2 1
(5.2.13)
G(y , x)
u
(x)dvg (x) + O
2
where C > 0 is independent of . Up to a subsequence, we can assume that

y y0 as +. Let us assume first that y0 = x0 . We fix > 0 such that
dg (x0 , y0 ) > 3. Then G(y , .) is bounded in Bx () so that
2 1
2 1
G(y , x)
u
(x)dvg (x) C
u
dvg .
Bx ()
Bx ()
Independently, we can write that
n+2
2 1
2
u
dvg
Bx (
)
n2
Bx (
)
dvg C
2 .
From (5.2.10), where we choose < 2/(2 1), we then get that
2 1
2 1
G(y , x)
u
(x)dvg (x) C
u
dvg
Bx ()
n+2

2 (2 1)
Bx (
)
+C
Bx ()\Bx (
)
It follows that
dg (x , y )
n2
Bx ()
dg (x , x)(+2n)(2
1)
dvg (x) .
n2
2 1
G(y , x)
u
(x)dvg (x) C
2 .
Coming back to (5.2.13), we then get that

dg (x , y )n2 u
(x )
u (y ) C
since < 2/(2 1). Now we assume that, up to a subsequence, y x0 as
for some C > 0 independent of . Then it
+, and that dg (x , y ) C
is clear that
dg (x , y )n2 u
(x )
u (y ) C u
(x )1 u
(y ) C .
78
CHAPTER 5
Finally, we assume that, up to a subsequence, y x0 as + and that

dg (x ,y )
+ as +. Coming back to (5.2.13), we write that
dg (x , y )n2 u
(x )
u (y )
n2
C
2 dg (x , y )n2
1
B
n2
2
+C
2 1
G(y , x)
u
(x)dvg (x)
dg (x , y )n2
2
B
2 1
G(y , x)
u
(x)dvg (x)

1)
n2
+O
2(2
d
(x
,
y
)
g

1
where B = x Bx () s.t. dg (y , x) 12 dg (x , y ) , and B2 is given by the
equation B2 = Bx ()\B1 . By a standard property of the Greens function (we
refer to Appendix A) there exists C > 0 such that G(x, y) Cdg (x, y)2n for all
(x, y) M M , x = y. We once again fix > 0 such that < 2/(2 1). Then,
n2
2 1
dg (x , y )n2 u
(x )
u (y ) C
2
u
dvg
n2
2
+C
1
B
dg (x , y )n2
2 1
dg (y , x)2n u
(x)dvg (x) + o(1) .

2
B
As above, using (5.2.10), we easily get that
n2
2 1
2
dvg C
1
B
for some C > 0 independent of . Independently, still using (5.2.10), we can write
that
n2
2
dg (x , y )n2

C

dg (x , y )
2 1
dg (y , x)2n u
(x)dvg (x)
2
B
2(2 1)
1
dg (x , y )2
2
B
dg (y , x)2n dvg (x)
2(2 1)
= o(1)
dg (x , y )
where C > 0 is independent of . Hence,
(x )
u (y ) C
dg (x , y )n2 u
where C > 0 is independent of . Summarizing these cases, we have
proved that,
up to a subsequence, the sequence dg (x , y )n2 u
(x )
u (y ) is bounded. In
particular, (5.2.9) is true.
C
From now on, we let G be the Greens function of the operator g + h , where
h is given by (5.0.1). Let also y0 M , and a sequence (y ) in M such that
y y0 as +.
S TEP 3. We claim that the two following asymptotics hold: if y0 = x0 , then, up
to a subsequence,

n2
2
2/n
n
n2
2
2
dg (x , y )
2 u
(y ) 1
(5.2.14)
4
79
as +, and if y0 = x0 , then, up to a subsequence,

n2
2
2/n
n
n2
2
2
+
dg (x , y )
2 u
(y )
4
(n 2)n1 dg (x0 , y0 )
n2
(5.2.15)
G(x0 , y0 )
as +. In particular, from standard elliptic theory,

n2

2
u
(x )
u (x) 4n2/n
(n 2)n1 G(x0 , x)
2
in Cloc
(M \{x0 }) as +. We prove (5.2.14) and (5.2.15). For large,
thanks to (5.0.1), g + h is coercive. We let G be the Greens function of the
operator g + h . Given > 0 and small, we write that
G (y , x) (g u
+ h u
) dvg (x)
u
(y ) =
M
2 1
=
G (y , x)
u
(x)dvg (x)
M
2 1
=
G (y , x)
u
(x)dvg (x)
Bx ()
+
From step 2,
2 1
G (y , x)
u
(x)dvg (x) .
M \Bx ()
2 1
G (y , x)
u
(x)dvg (x)
M \Bx ()
n+2
2
=O
M \Bx ()
G (y , x)dvg (x)
n+2
=O
2
.
Therefore,
u
(y ) =
Bx ()
n+2
2 1
.
G (y , x)
u
(x)dvg (x) + O
2
(5.2.16)
We again distinguish three cases. First we assume that y0 = x0 . We choose > 0

1
and small such that dg (x0 , y0 ) 3. For x B0 (
), the Euclidean ball of
1
center 0 and radius
, we let
n2

2 u
expx (
x)
u
(x) =
where expx is the exponential map at x . Then,
2 1
u
g u
+ h
= u

(x) =
where h
2 h expx (
x) and g (x) = (expx ) g(
x). The u
s are
0 uniformly. It follows from standard
clearly bounded, and, from (5.0.1), h
80
CHAPTER 5
0,
Cloc
(Rn )
elliptic theory that the u

s are bounded in
for some > 0, and then,
from Ascolis theorem, that, up to a subsequence, we also have that u
u in
0
n
(R
)
as
+.
Moreover,
it
is
easily
checked
that
u(0)
=
1
=
maxRn u,
Cloc

that Rn u2 dx = 1, and that
u =
1 2 1
u
Kn2
where is the Euclidean Laplacian. Hence,

u(x) =
2/n
n
|x|2
1+
4
n2
2
where n is the volume of the unit n-sphere. Coming back to (5.2.16) we write
that
u (y )
u
(x )
n2
=
2
Bx ()
B0 (
1
)
2
2 1
G (y , x)
u
(x)dvg (x) + O
2 1
2

G y , expx (
x) u
(x)dvg (x) + O
.
An equivalent formulation of the claim in step 2 is that u

Cu for some C > 0
independent of , where u is as above. The Lebesgue-dominated convergence
theorem and this estimate then give that

2
lim u
(x )
u (y ) = Kn G(y0 , x0 )
u2 1 dx
+
Rn
where G is the Greens function of the operator g + h . It follows that

u
(x )
u (y )
4
2/n
n2
2
(n 2)n1 G(x0 , y0 )
as +. This proves (5.2.15). Let us now assume that y0 = x0 and that

dg (x ,y )
R as +. We let z be such that y = expx (
z ). Then

lim dg (x , y )
n2
u
(x )
u (y ) = lim
dg (x , y )
n2
u
(z )
u
where u
is as above. Clearly, |z | R as +, and since, as above, u
dg (x ,y )
0
n
in Cloc (R ) as +, we get that if y0 = x0 and
R as +,
then
n2

2
2
R
(x )
u (y )
(5.2.17)
dg (x , y )n2 u
2/n
1 + n4 R2
81
as +. Finally, we assume that y0 = x0 and that

+. Coming back to (5.2.16), we can write that
dg (x , y )n2 u
(x )
u (y )
n2
=
2 dg (x , y )n2
n2
2
2 1
(x)dvg (x)
G (y , x)
u
dg (x , y )n2
+ as
1
B
dg (x ,y )
2 1
G (y , x)
u
(x)dvg (x)
2
B

2
+O dg (x , y )n2

where B1 = x Bx () s.t. dg (y , x) 12 dg (x , y ) , and B2 is given by the
equation B2 = Bx ()\B1 . Clearly,
n2
2 1
G (y , x)
u
(x)dvg (x)
1
B
2 1

G y , expx (
x) u
(x)dvg (x)
1
1
where =
1
x). For x , the ratio
expx (B ). Let z = expx (
dg (y ,z )
+ as +. Writing that
dg (y , z ) dg (x , z ) dg (x , y ) dg (y , z ) + dg (x , z )
and noting that dg (x , z ) =
|x|, we get that
lim
dg (x , y )
= 1.
dg (y , z )
By a standard property of the Greens function (we refer to Appendix A for more
details) this implies that

1
lim dg (x , y )n2 G y , expx (
x) =
+
(n 2)n1
where n1 is the volume of the unit (n 1)-sphere. Because of the Lebesguedominated convergence theorem and step 2, we then get that
n2
2 1
2 dg (x , y )n2
G (y , x)
u
(x)dvg (x)
lim
+
1
=
(n 2)n1 Kn2
1
B
2 1
dx =
n2
2
2/n
n
where u is as above. On the other hand, from step 2 [see (5.2.9)] we can write that
n2
2 1
2 dg (x , y )n2
G (y , x)
u
(x)dvg (x)

C

C
dg (x , y )
dg (x , y )
Rn
2
2
2
B
1
dg (x , y )2
= o(1)
2
B
dg (y , x)2n dvg (x)
82
CHAPTER 5
dg (x ,y )
where C > 0 is independent of . Therefore, if y0 = x0 and

+, then

n2
2
4
dg (x , y )n2 u
(x )
u (y )
2/n
n
+ as
(5.2.18)
as +. Combining (5.2.17) and (5.2.18), and since u(0) = 1, this proves

(5.2.14).
Now we prove (5.2.7) and (5.2.8), and thus Theorem 5.2. We let v be the
function defined by
n2

2
u
(x)
v (x) =
2/n
2 + n4 dg (x , x)2
and let y be such that v (y ) = minM v . Then, up to a subsequence, it follows
from (5.2.14) and (5.2.15) that v (y ) c for some c > 0 independent of .
Together with step 2, this proves (5.2.7). Independently,
lim dg (x0 , y)n2 G(x0 , y) =
yx0
1
.
(n 2)n1
Hence, given > 0, there exists > 0 such that if y Bx0 ( ), then
1
(n 2)n1 dg (x0 , y)n2 G(x0 , y) 1 + .
1+
If we let y and y be such that
y ) = max v ,
v (y ) = min v and v (
Bx0 ( )
Bx0 ( )
we then get with (5.2.14) and (5.2.15) that, up to a subsequence,

1
v (y ) v (
y ) 1 + .
1+
This proves (5.2.8) and thus Theorem 5.2.
Chapter Six
Asymptotics When the Energy Is Arbitrary
g u + h u = u2
(E )

so that u > 0 and
2 1
g u + h u = u
for all . We describe in this chapter the C 0 -theory for the blow-up behavior of the
u s when the energy of the u s is arbitrary.
As in Chapters 4 and 5, we assume in what follows that there exist 0 < < 1
and a smooth (or only C 0, ) function h on M such that
(6.0.1)
h h in C 0, (M ) as + .
We also assume that there exists > 0 such that E(u ) for all , where
E(u) = u2 is as in Chapter 2. The existence of u implies the coercivity of
the operator g + h . An estimate like the one in Theorem 6.1 below implies the
coercivity of the operator g + h (see Appendix B). Independently, as in Chapter
4, multiplying (E ) by u and integrating over M , we get that

2
2
|u | dvg +
h u dvg =
u2 dvg 2 .
M
Noting that by Holders inequalities,

2/n

n/2
2

h
u
dv
|h
|
dv
g
g

M
and that by (6.0.1),
u2 dvg
2/2
|h |n/2 dvg C
M
for all , where C > 0 is independent of , we get that the u s are bounded in
H12 (M ). Up to a subsequence, we may therefore assume that for some function
u0 H12 (M ),
u u0 weakly in H12 (M )
0
as +. It is easily seen that u is a solution of

g u0 + h u0 = (u0 )2
(6.0.2)
84
CHAPTER 6
By the maximum principle and regularity theory, u0 C 2, (M ), and either u0 0

or u0 > 0 everywhere in M . We assume in what follows that
lim max u = +
(6.0.3)
+ M
so that blow-up occurs. If this is not the case, up to a subsequence, we easily get
from standard regularity theory that u u0 strongly in C 2 (M ) as +,
and this provides a complete description of the behavior of the u s. We prove
in this chapter that the following theorem holds. The theorem was announced in
Druet-Hebey-Robert [32]. Applications of the theorem are in Druet [27].
T HEOREM 6.1 Let (M, g) be a smooth compact Riemannian manifold of dimension n 3, (h ) be a sequence of smooth functions on M such that (6.0.1) is satisfied and (u ) be a sequence of positive solutions to (E ) such that E(u )
for some > 0 and all . We assume that (6.0.3) holds. Then there exist N N ,
converging sequences (xi, ) in M , and sequences (i, ) of positive real numbers
converging to 0, i = 1, . . . , N , such that, up to a subsequence,
n2
2
N
1
i,
0
(1 ) u (x) +
C i=1 2 + dg (xi, ,x)2
i,
u (x) (1 + ) u0 (x) + C
n(n2)
N
i=1
i,
2i,
dg (xi, ,x)2
n(n2)
n2
2
for all x M and all , where u is given by (6.0.2), C > 1 is independent

of and x, and ( ), independent of x, is a sequence of positive real numbers
converging to 0 as +. In particular, the u s are C 0 -controlled, on both
sides, by u0 and standard bubbles.
A complement to Theorem 6.1 is that C can be chosen as close as we want to 1 if
we restrict the equation in Theorem 6.1 to small neighborhoods of the geometrical
blow-up points. For instance, if u0 0, or if the u s just have one geometrical
blow-up point, then for any > 0, there exists > 0 such that, up to a subsequence,
n2
2
N

1
i,
0
(1 ) u (x) +
1 + i=1 2 + dg (xi, ,x)2
i,
n(n2)
u (x) (1 + ) u0 (x) + (1 + )
N

i=1
i,
2i, +
dg (xi, ,x)2
n(n2)
n2
2
for all , all x0 S, and all x Bx0 ( ), where S is the set consisting of the
limits of the xi, s as +. Outside the Bx ( )s, x S, the u s converge
C 2, to u0 . The estimate then extends to M in the particular case where u0 0.
We refer to section 6.3 [see, in particular, (6.3.24), (6.3.25), and (6.3.37)] for more
details and a refined statement.
ASYMPTOTICS WHEN THE ENERGY IS ARBITRARY
85
Another complement to Theorem 6.1 is that the bubbles in this theorem satisfy
Theorem 3.1. More precisely, we also have that for any x M and any ,
n2
2
N

i,
0
u (x) = u (x) +
+ R (x)
d (xi, ,x)2
2i, + gn(n2)
i=1
where the R s, R H12 (M ) for all , are such that R 0 strongly in H12 (M )
as +. Moreover, if

1
1
|u|2 dvg
|u|2 + h u2 dvg
Ig (u) =
2 M
2 M
and

1
1
Ig (u) =
|u|2 dvg
|u|2 + h u2 dvg
2 M
2 M
are as in Chapter 3, then
N
Ig (u ) = Ig (u0 ) + Knn + r
n
where the r s, r R for all , are such that r 0 as +. We refer to
the end of section 6.3 for the proof of these claims.
Sections 6.16.3 below are devoted to the proof of Theorem 6.1. We set up
some notation in the rest of this introduction. We let the xi, s and i, s be as
in Theorem 4.1, i = 1, . . . , N , and choose N to be maximal. As a remark, the
xi, s in Theorem 6.1 will be slightly distinct from those in Theorem 4.1, but the
i, s are the same. Up to a subsequence, we may assume that for all i, j, k in
{1, . . . , N }, the 1
k, dg (xi, , xj, )s have a limit in [0, +] as +. We let
S = {xi, , i = 1, . . . , N }
(6.0.4)
and rearrange the xi, s in the following way. First, we let x1, be such that 1,
satisfies the condition 1, i, for all i = 1, . . . , N . We let N1 = 1 and
S1 = {xN1 +1, , . . . , xN2 1, }
be the subset of S consisting of the xi, s, i > N1 , which are such that
lim
dg (xN1 , , xi, )
< + .
N1 ,
We let then xN2 , be such that N2 , i, for all i = N2 , . . . , N , and let

S2 = {xN2 +1, , . . . , xN3 1, }
be the subset of S consisting of the xi, s, i > N2 , which are such that
lim
dg (xN2 , , xi, )
< + .
N2 ,
We repeat the process up to the exhaustion of S . We then get the existence of

some integer p such that S consists of the (ordered) sequence
x1, = xN1 , , x2, , . . . , xN2 , , xN2 +1, , . . . , xN3 , ,
. . . , xNp , , . . . , xN, = xNp+1 1, .
86
CHAPTER 6
Given i = 1, . . . , p, we let
S0 = {xNi , , i = 1, . . . , p} ,

Si = xNi +1, , . . . , xNi+1 1, ,
S = p S i .
(6.0.5)
i=1
The
Sj s,
j = 0, . . . , p, are disjoint, and we have that

S = S0 S .
As a remark, it might be, in this process of rearranging the xi, s, that Ni+1 =
Ni + 1 or, equivalently, that Si = . In this case, xNi+1 , is chosen such that
Ni+1 , i, for all i = Ni + 1, . . . , N .
Given i = 1, . . . , p, we let
yi, = xNi , and i, = Ni , .
(6.0.6)
The above construction gives that

1, p,
(6.0.7)
dg (yi, , yj, )
+
i,
(6.0.8)
and that for i = j,
as +. We let R0 > 0 be such that for any j = 1, . . . , p and for any

i = Nj , . . . , Nj+1 1,
lim
dg (yj, , xi, )
R0
.
j,
2
(6.0.9)
Independently, we let R be the function given by

R (x) = min dg (yj, , x) .
j=1,...,p
(6.0.10)
Up to a subsequence, we may assume that there exists C > 0 such that, for any ,
sup
and that
lim
lim
sup
R+ + M \0 (R)
R (x) 2 1 u (x) C ,
n
M \0 (R0 )

n
R (x) 2 1 u (x) u0 (x) = 0
(6.0.11)
(6.0.12)
where 0 (R) = pj=1 Byj, (Rj, ). We prove (6.0.11) and (6.0.12) in what follows. Let (y ) be a sequence of points in M \0 (R0 ). We distinguish two cases.
In the first case we assume that there exists j {1, . . . , p} such that
dg (yj, , y )
R
j,
as +, where R R0 . By the above construction,
R (y ) = dg (yj, , y )
87
for large. We write that y = expyj, (j, z ), where z Rn is such that

|z | R as +. From Theorem 4.1, proposition (P1), and (6.0.8), (6.0.9),
there exists z0 Rn such that, if z z as +, then
n
n
lim R (y ) 2 1 u (y ) = R 2 1 u (z z0 )
+
where u is given by (4.1.5). In particular,

n
lim sup R (y ) 2 1 u (y ) < + .
+
In the second case we assume that, for any j {1, . . . , p},

dg (yj, , y )
+
j,
N
be as in Theorem 4.1. After passing to a subsequence,
as +. We let R
there exists k {1, . . . , N } such that
N
R
(y ) = dg (xk, , y ) .
We let j {1, . . . , p} be such that Nj k Nj+1 1. By (6.0.9), and for
sufficiently large,
N
R
(y ) dg (y , yj, ) dg (xk, , yj, )
dg (y , yj, ) R0 j,
1
dg (y , yj, )
2
1
R (y ) .
2
By the first claim of proposition (P2) of Theorem 4.1, it follows that, here again,
n
lim sup R (y ) 2 1 u (y ) < + .
+
In particular, (6.0.11) is proved. Independently, it is easily checked that, for any

i {1, . . . , p} and for any l {Ni , . . . , Ni+1 1},

dg (xl, , y )
i, dg (yi, , y )
R0
l,
l,
i,
so that
dg (xl, , y )
+
l,
as +. In particular, for any R > 0, y M \ (R) for large, where
(R) is as in Theorem 4.1. By the second claim of proposition (P2) of Theorem
4.1, and since
1
N
R
(y ) R (y ) ,
2
we get that

n
R (y ) 2 1 u (y ) u0 (y ) 0
as +. This proves (6.0.12).
The rest of this chapter is divided as follows. Section 6.1 is devoted to the proof
of an upper estimate on the u s, similar to the one in Theorem 6.1 when we replace
S by S0 . In section 6.2 we prove the upper estimate on the u s with respect to
the full set S . Section 6.3 is devoted to the proof of the estimate from below in
Theorem 6.1, and to the asymptotics of u .
88
CHAPTER 6
6.1 A FUNDAMENTAL ESTIMATE: 1

This section is devoted to the proof of a fundamental pointwise estimate with respect to S0 , where S0 is as in (6.0.5). We borrow ideas from Druet-Robert [33]
and Hebey-Vaugon [47, 48] where this kind of analysis was carried out in the case
when the energy of the u s is of minimal type (see Chapter 5). We divide this section into two subsections. In subsection 6.1.1 we consider the case where u0 0.
In subsection 6.1.2 we consider the case where u0 0. All that follows is up to a
subsequence.
6.1.1 The case u0 0. We assume in this subsection that u0 0. Let R0
be as in (6.0.9). We want to prove that there exists C > 0 such that for any
y M \ pj=1 Byj, (R0 j, ),
u (y) C
p
2n
2
j,
dg (yj, , y)
(6.1.1)
j=1
where the yj, s and j, s are as in (6.0.6). We split the proof of (6.1.1) into
several steps that we refer to as claims. The proof of (6.1.1) is by induction.
First we introduce some notation. We let h0 C 0, (M ) be such that h0 < h
and g + h0 is coercive. We let G : M M \ {(x, x) , x M } R+ , be the
Greens function of the operator g + h0 . Then G is the only symmetric positive
function satisfying in the sense of distributions
g,y G (x, y) + h0 (y)G (x, y) = x
for all x M , where x is the Dirac mass at x. We know (see Appendix A) that
there exist > 0, C1 > 0 and C2 > 0 such that, for any x = y in M ,
dg (x, y) =
C1
|y G(x, y)|
G(x, y)
dg (x, y)
(6.1.2)
and
1
n2
dg (x, y)
G (x, y) C2 .
C2
(6.1.3)
Our first claim is the following:

C LAIM 6.1.1. Assume that u0 0, and let 0 < < 12 . If is sufficientl small,
then there exist R() > 0 and C () > 0 such that for any > 0 and any y in
M \ pj=1 Byj, (R () j, ),
(1 n )(12)
(n2)(1)
1, 2
R (y)
u (y) C ()
where R is as in (6.0.10), the j, s are as in (6.0.6), and the yj, s are as in
(6.0.6).
Proof of claim 6.1.1. Let L be the linear operator given by
2 2
u.
L (u) = g u + h u u
89
Since L u = 0 and u > 0, L satisfies the maximum principle on M (see for

instance [9]). We let 0 < < 12 be such that h > (1 ) h0 and set
H (y) =
p
Gj, (y)
j=1
where
Gj, (y) = G (yj, , y)
and G is the Greens function of g + h0 . Easy computations give that for any
y M \ {y1, , . . . , yp, },
L H (y) =
p
Gj, (y)
A (y)
(6.1.4)
j=1
where
2 2
A (y) = h (y) (1 ) h0 (y) u (y)
|Gj, (y)|2
+ (1 )
Gj, (y)
We claim that L H 0 in M \ pj=1 Byj, (Rj, ) for some R > 0 to be fixed

later on, and for large enough. In order to prove this claim, we let (y ) be a
sequence of points in M \ pj=1 Byj, (Rj, ). Up to a subsequence, y y0 as
+. We distinguish two cases.
Case 1. We assume that dg (y0 , S) 2 , where S is given by Theorem 4.1 and
0
(M \S). Coming
is as in (6.1.2). By point (P1) of Theorem 4.1, u 0 in Cloc
back to (6.1.4), we can write in this case that, for sufficiently large,
L H (y )
p
Gj, (y )
(h (y ) (1 ) h0 (y ) + o (1))
j=1
0
since we fixed such that h > (1 ) h0 .
Case 2. We assume that dg (y0 , S) 2 , where S is given by Theorem 4.1
and is as in (6.1.2). We let i {1, . . . , p} be such that, up to a subsequence,
R (y ) = dg (yi, , y ). We write that
1
L H (y ) (1 ) Gi, (y )
2 2
u (y )
p
|Gi, (y )|2
2
Gi, (y )
1
Gj, (y )
j=1
Then, from (6.1.2) and (6.1.3),

C12
(2n)(1)2
R (y )
C21
p

2 2 1
(2n)(1)
u (y )
C2
dg (yj, , y )
,
L H (y ) (1 )
j=1
90
CHAPTER 6
and we get that

C12
(2n)(1)2
R (y )
C21
L H (y ) (1 )
2 2
u (y )
(2n)(1)2
(2n)(1)
C21 pR (y )

C12
2
2 2
1
(1 ) 1 pC2 R (y ) u (y )
C2
R (y )
for sufficiently large. We fix R > R0 , depending only on , such that, for
large,
2
2 2
R (y ) u (y )
(1 ) C12
.
2p
C222
Since u0 0, this is always possible from (6.0.12). Then L H (y ) 0.

In particular, it follows from cases 1 and 2 that there exists R () R0 such that,
for large,
L H 0 in M \ pj=1 Byj, (R () j, ) .
(6.1.5)
Since R () R0 , we can write that, for large,

1 n
u (y) 2j, 2
(6.1.6)
for any j {1, . . . , p} and any y Byj, (R () j, ). This follows from point
(P1) of Theorem 4.1 and the choice of R0 we made in (6.0.9). Independently,
(6.1.3) implies that, for any j {1, . . . , p},
(2n)(1)
H C21 (R () j, )
(6.1.7)
on Byj, (R () j, ). Using (6.1.6) and (6.1.7), and since 1, j, for all

j {1, . . . , p}, we get that for any j {1, . . . , p} and any y Byj, (R () j, ),
1 n
u (y) 2j, 2
( n 1)(12) (2n)(1)
2j,2
j,
(n2)(1)
2C21 R ()
( n2 1)(12)
1,
H (y) .
In particular,
( n2 1)(12)
H (y)
u (y) C()1,

for all j {1, . . . , p} and all y Byj, R () j, . Noting that L u = 0, it
follows from (6.1.5) that

( n2 1)(12)
H u 0
L C()1,

in M \pj=1 Byj, R () j, . From the maximum principle, we then get that there
exists C() > 0, depending only on , such that
( n2 1)(12)
H (y)
u (y) C () 1,
91
in M \
pj=1
Byj, (R () j, ). Together with (6.1.3), this proves claim 6.1.1. 2
The content of our second claim is that the inequality in claim 6.1.1 holds also
with = 0. Claim 6.1.2 can be stated as follows:
C LAIM 6.1.2. Assume that u0 0. There exists C > 0 such that for any > 0
and any y in M \ pj=1 Byj, (R0 j, ),
1
2n
2
u (y) C1,
R (y)

(6.0.6).
Proof of claim 6.1.2. For j {1, . . . , p}, we let
j, = {y M s.t. R (y) = dg (yj, , y)} .
We also let G be the Greens function of g + h0 , where h0 C 0, (M ) is such
that g + h0 is coercive and h > h0 . Let (y ) be a sequence in M \ pj=1
Byj, (R0 j, ). We have to prove that
1 n
n2
lim sup 1, 2 R (y )
+
u (y ) < + .
(6.1.8)
By the Greens representation formula, we get, using (E ) and h > h0 , that
2 1
u (y )
G (y , y) u (y)
dvg
(6.1.9)
M
for large. We distinguish three cases.

Case 1. We assume that, up to a subsequence, R (y ) 2 as + for
some > 0. We then write that, for any j {1, . . . , p},
2 1
G (y , y) u (y)
dvg
j,
2 1
j, Byj, ()
G (y , y) u (y)
dvg
2 1
j, \Byj, ()
G (y , y) u (y)
dvg .
By (6.1.3), G (y , y) stays bounded in Byj, () since dg (yj, , y )

large. Thus we can write that
2 1
G (y , y) u (y)
dvg
j, Byj, ()
j, Byj, ()
2 1
u (y)
(6.1.10)
3
2
for
(6.1.11)
dvg
92
CHAPTER 6
where C, as in the following, is a positive constant independent of . By claim

6.1.1, for any > 0, there exist C () and R () such that
2 1
u (y)
dvg
j, Byj, ()
2 1
j, Byj, (R()1, )
(n+2)(12)
2
u (y)
dvg
+C () 1,
j, Byj, ()\Byj, (R()1, )
dg (yj, , y)
(n+2)(1)
dvg .
Here we used that j, Byi, (R () i, ) = for large enough as soon as

2
, and using Holders inequality,
i = j. This follows from (6.0.8). Taking < n+2
we get that
2 1
u (y)
dvg
j, Byj, ()
1

Volg Byj, (R () 1, ) 2 u 22 1
(n+2)(12)
(n+2)(1)
2
+C () 1,
dg (yj, , y)
dvg

n2
2
M \Byj, (R()1, )
(n+2)(12)
2
= O 1, + O 1,
n2
= O 1,2
.
n(n+2)(1)
1,
Coming back to (6.1.11), we have obtained that
n2
2 1
.
G (y , y) u (y)
dvg = O 1,2
j, Byj, ()
(6.1.12)
On the other hand, using claim 6.1.1 and (6.1.3), we can write that, for any > 0,
2 1
G (y , y) u (y)
dvg
j, \Byj, ()
( n+2
2 )(12)
C1,

( n+2
2 )(12)
= O 1,
Taking <
2
n+2 ,
j, \Byj, ()
G (y , y) dvg (y)
this implies that

2 1
j, \Byj, ()
G (y , y) u (y)
n
2 1
.
dvg = O 1,
(6.1.13)
Coming back to (6.1.10), we get with (6.1.12) and (6.1.13) that for any j, with
j {1, . . . , p},
n
2 1
2 1
.
G (y , y) u (y)
dvg = O 1,
j,
93
Noting that M =
that (6.1.8) holds.
pj=1 j, ,
and using (6.1.9), we have proved that (y ) is such
Case 2. We assume that there exists j {1, . . . , p} such that, up to a subsequence,

dg (y , yj, )
R
j,
as +, where R R0 . In this case, it easily follows from point (P1) of
Theorem 4.1 and the definition (6.0.9) of R0 that (y ) is such that (6.1.8) holds.
Just note that 1, j, for any j {1, . . . , p}.
Case 3. We assume that R (y ) 0 as + and that for any j with
j {1, . . . , p}, up to a subsequence,
dg (yj, , y )
+
j,
as +. Given k {1, . . . , p}, we let

1
1
= y k, Byk, () s.t. dg (y , y) R (y ) ,
Bk,
2

1
2
= y k, Byk, () s.t. dg (y , y) < R (y ) .
Bk,
2
We write that
2 1
G (y , y) u (y)
dvg
k,
2 1
1
Bk,
G (y , y) u (y)
2
Bk,
dvg
2 1
G (y , y) u (y)
(6.1.14)
dvg
2 1
k, \Byk, ()
G (y , y) u (y)
dvg .
As in case 1 [see the proof of (6.1.13)] we can write that
n
2 1
2 1
.
G (y , y) u (y)
dvg = O 1,
(6.1.15)
k, \Byk, ()
Independently, from (6.1.3),

2n
G (y , y) C2 dg (y , y)
2n
2n2 C2 R (y )
1
in Bk,
, so that
2 1
2n
G (y , y) u (y)
dvg CR (y )
1
Bk,
1
Bk,
We proved in case 1 that
k, Byk, ()
2 1
u (y)
2 1
u (y)
n
2 1
.
dvg = O 1,
dvg .
94
CHAPTER 6
It follows that

n
2 1
2n
2 1
.
G (y , y) u (y)
dvg = O 1,
R (y )
(6.1.16)
1
Bk,
2
, we have that
In Bk,
dg (yk, , y) dg (yk, , y ) dg (y , y)
1
R (y ) R (y )
2
1
R (y ) .
2
Let > 0, to be fixed later on. Since
dg (yk, , y )
+
k,
2
as +, we have that dg (yk, , y) R () k, for all y Bk,
and large.
Noting that for large and i = k,
k, Byi, (R () i, ) = ,
2
we can apply claim 6.1.1. It follows that, for any y Bk,
,
( n2 )(12)
(n2)(1)
dg (yk, , y)
.
u (y) C1,2
Together with (6.1.3) we then get that
2 1
G (y , y) u (y)
dvg
2
Bk,
( n+2 )(12)
(n+2)(1)
C1,2
R (y )
( n+2
2 )(12)
C1,
(n+2)(1)
R (y )
2
Bk,
G (y , y) dvg
By (
2n
R (y )
2
dg (y , y)
dvg .
This easily leads to
( n+2 )(12)
2 1
2(n+2)(1)
R (y )
G (y , y) u (y)
dvg C1,2
2
Bk,
n
2 1
2n
C1, R (y )
R (y )
1,
(n+2)2
.
1
Let us assume that 1,
R (y ) + as +. We let > 0 be such that
2
< n+2 . Then,

n
2 1
2n
2 1
.
(6.1.17)
G (y , y) u (y)
dvg = O 1,
R (y )
2
Bk,
Since M = pk=1 k, , we get, by combining (6.1.9) with (6.1.14)(6.1.17), that

1
R (y ) = O (1), we get as
(y ) is such that (6.1.8) holds. If, on the contrary, 1,
a straightforward consequence of (6.0.11) that (y ) is such that (6.1.8) holds also.
95
From cases 13, if (y ) is a sequence of points in

1 n
2
n2
lim sup 1, R (y )
+
M \ pj=1 Byj,
(R0 j, ), then
u (y ) < + .
2
This ends the proof of claim 6.1.2.

Given j {1, . . . , p} and > 0, we let
Rj, (y) = min dg (yk, , y) and
k=j,...,p
( n 1)(12)
(2n)(1)
j, (y) = j,2
dg (yj, , y)
.
(6.1.18)
Then we say that (6.1.19)j holds if there exists C > 0, independent of , such that
for any > 0 and any y M \ pi=1 Byi, (R0 i, ),

j1
n 1
n
2n
2n
2
2 1
+ j, Rj, (y)
.
(6.1.19)j
i, dg (yi, , y)
u (y) C
i=1
By claim 6.1.2, (6.1.19)j holds for j = 1. We prove the following:

C LAIM 6.1.3. Assume that u0 0. Let 0 < < 12 , j {2, . . . , p}, and assume
that (6.1.19)j1 holds. If is sufficientl small, then there exist R () > R0 and
C () > 0, such that for any > 0 and any y M \ pi=1 Byi, (R () i, ),

j1

( n2 1)(12)
(2n)(1)
i, (y) + j,
Rj, (y)
u (y) C ()
i=1
where the
are as in (6.1.18), the Rj, s are as in (6.1.18), the j, s are as
in (6.0.6), and the yj, s are as in (6.0.6).
i, s
Proof of claim 6.1.3. We let 0 < <

h
1
2
sufficiently small such that
= h (1 ) h0 > 0
(6.1.20)
where h0 is as above. Given k {1, . . . , j 1}, we let

k, = y M s.t. k, (y) i, (y) for all i = 1, . . . , j 1 . (6.1.21)
It is easily checked that j1

i=1 i, = M . We also let

(1 ) C12
1
; 2 inf h
D () = min
2(1) 2
M
2p
C
(6.1.22)
where C1 , C2 , and are given by (6.1.2) and (6.1.3), and where h is given by
(6.1.20). As a starting point
we claim

that there exists R () > R0 such that for
any y M \ pi=1 Byi, R () i, and any k {1, . . . , j 1},
2
2 2
y k, = min {Rj, (y) , dg (yk, , y)} u (y)
D ()
(6.1.23)
for large. We prove this claim using

(6.1.19)
j1 and (6.0.12). Let (y ) be a
sequence of points in M \ pi=1 Byi, Ri, , R > R0 to be chosen later on, such
96
CHAPTER 6
that y k, for some k {1, . . . , j 1}. We need to prove that, up to choosing

R sufficiently large,
2
2 2
min {Rj, (y ) , dg (yk, , y )} u (y )
D ()
(6.1.24)
for large. From (6.1.19)j1 , and since

n
2n
2
i,
dg (yi, , y )
we have that
u (y ) C
j2
(n2)
= i,
(2n)
dg (yi, , y )
i, ,

n
2 1
2n
i, dg (yi, , y )
n
2 1
2n
+ j1, Rj1, (y )
i=1
j2

R
(2n)
i,
n
2 1
2n
(y ) + j1, Rj1, (y )
i=1
Noting that
Rj1, (y ) = min {Rj, (y ) , dg (yj1, , y )} ,
we easily get that
2n
Rj1, (y )
2n
Rj, (y )
2n
+ dg (yj1, , y )
Since y k, , it follows that

j1

n
2n
(2n)
2 1
R
i, (y ) + j1, Rj, (y )
u (y ) C
!
i=1
n
2n
2
Rj, (y )
C (j 1) R(2n) k, (y ) + j1,
Therefore
n 1

min {Rj, (y ) , dg (yk, , y )} 2 u (y )

C (j 1) R
(2n)
dg (yk, , y )

C (j 1) R
1 n
2

+
Rj, (y )
j1,
n
2 1
k,
(y ) +
1 n2
Rj, (y )
j1,
"
.
1 n2
Assuming that
Rj, (y )
+
j1,
as +, and choosing R sufficiently large, it easily follows from the above
equation that (6.1.24) holds. Let us assume now that
Rj, (y ) = O (j1, ) .
(6.1.25)
We claim that in such a case Rj, (y ) = R (y ). Then (6.1.24) follows from

(6.0.12) by choosing R sufficiently large. To see that Rj, (y ) = R (y ), we let
97
l {j, . . . , p} be such that Rj, (y ) = dg (y , yl, ). Given i {1, . . . , j 1},

from (6.1.25), we have that
dg (y , yi, ) dg (yi, , yl, ) dg (y , yl, )
dg (yi, , yl, ) O (j1, )

dg (yi, , yl, )
j1,
O (1)
j1,

dg (yi, , yl, ) i,
j1,
O (1) .
i,
j1,
1
+ as + by (6.0.8) and j1, i, by
Since dg (yi, , yl, ) i,
(6.0.7), we get that
1
+
dg (y , yi, ) j1,
as +. From (6.1.25) we then get that, for sufficiently large,

dg (y , yi, ) Rj, (y )
for all i {1, . . . , j 1}. In particular, Rj, (y ) = R (y ). As already mentioned, this proves (6.1.24) and thus (6.1.23).
We now let Aj, be given by
( n2 1)(12)
Aj,
=
where
sup
sup
i=1,...,j1 yi,
i, (y)
( n 1)(12)
(2n)(1)
j,2
Rj, (y)

i, = y i, \ pk=1 Byk, (R () k, )
2
2 2
s.t. dg (yi, , y) u (y)

> D () .
(6.1.26)
(6.1.27)
By convention, Aj, = if the i, s, i = 1, . . . , j 1, are empty. We claim

that, for R () sufficiently large,
lim sup Aj,
+
j,
1.
j1,
(6.1.28)
We prove (6.1.28) in what follows. We assume that Aj, = . Then, up to a

subsequence, there exist k {1, . . . , j 1} and y k, such that
k, (y ) = (Aj, j, )( 2
1)(12)
(2n)(1)
Rj, (y )
We proceed by contradiction. Assuming that (6.1.28) is false, we have that

( n2 1)(12)
(2n)(1)
k, (y ) j1,
Rj, (y )
.
(6.1.29)
98
CHAPTER 6
From (6.1.19)j1 , and since y k, , this implies that

D ()
n2
4
dg (yk, , y ) 2
u (y )
j1
!

n
1
(n2)
Cdg (yk, , y ) 2
i, (y )
R ()
n
2 1
2n
"
i=1
+j1, Rj, (y )
!
n
1
(n2)
k, (y )
(j 1) R ()
Cdg (yk, , y ) 2
(n2) "

j1,
( n2 1)(12)
(2n)(1)
Rj, (y )
+j1,
Rj, (y )
!
n
(n2)
2 1
(j 1) R ()
k, (y )
Cdg (yk, , y )
(n2) "

j1,
+k, (y )
Rj, (y )
(n2) "

!
j1,
1 n
1 n
(
2
2 )(12)
.
+ R ()
C (j 1) R ()
Rj, (y )
Taking R () sufficiently large, this implies in turn that
Rj, (y ) = O (j1, ) .
(6.1.30)
Coming back to (6.1.29), we get with (6.1.30) that

n 1 (12)
k, ( 2 )
(n2)(1)
(n2)(1)
dg (yk, , y )
Rj, (y )
j1,

n
( 2 1)(12) n2 1
= O k,
j1, .
Since k j 1, j1, k, . Thus,
dg (yk, , y ) = O (k, ) .
(6.1.31)
We let l {j, . . . , p} be such that Rj, (y ) = dg (yl, , y ). From (6.1.30) and

(6.1.31) we can write that
dg (yk, , yl, ) dg (yk, , y ) + dg (yl, , y )
= O (k, ) + O (j1, )
= O (k, ) ,
a contradiction with (6.0.8). This proves (6.1.28).
Let G be the Greens function of g + h0 . We let
j, = j, max {Aj, ; 1}
(6.1.32)
and
H (y) =
j1
p

( n 1)(12)
( n 1)(12)
1
1
i,2
Gi, (y)
+ j,2
Gi, (y)
i=1
i=j
99
where Gi, (y) = G (yi, , y). It is clear from point (P1) of Theorem 4.1, (6.0.7),
(6.0.9), (6.1.3), and (6.1.32) that there exists C () > 0 such that

u C () H on Byi, R () i,
(6.1.33)
for all i {1, . . . , p}. We let L be the linear operator given by
2 2
L (u) = g u + h u u
u.
Straightforward computations give that

j1
!

( n 1)(12)
1
h (y) (1 ) h0 (y)
L H (y) =
i,2
Gi, (y)
i=1
2 2
u (y)
+ (1 )
|Gi, (y) |2 "

2
Gi, (y)
p
!

n
( 1)(12)
1
+ j,2
h (y) (1 ) h0 (y)
Gi, (y)
i=j
2 2
u (y)
+ (1 )
(6.1.34)
|Gi, (y) |2 "

2
Gi, (y)
for all y M \ pi=1 Byi, (R () i, ). We claim that, for sufficiently large,

L H 0 in M \ pi=1 Byi, (R () i, ) .
In order to prove this claim, we let (y ) be a sequence of points
y M \ pi=1 Byi, (R () i, ) .
We assume that y k, for some k {1, . . . , j 1} and distinguish four cases.
Case 1. We assume that
2 2
u (y )
D ()
2
where D () is as in (6.1.22) and is given by (6.1.2). Then

2 2
h (y ) (1 ) h0 (y ) u (y )
>0
for large. Thus L H (y ) 0 for large.

2 2
u (y )
D ()
,
2
that y k, , and that dg (yk, , y ) Rj, (y ). Since y k, ,

2
dg (yk, , y )
2
2 2
D () dg (yk, , y ) u (y )
D () .
2
Hence,
Rj, (y ) dg (yk, , y ) .
(6.1.35)
100
CHAPTER 6
We let l {j, . . . , p} be such that Rj, (y ) = dg (yl, , y ). By (6.1.35), using

(6.1.2) and (6.1.3), we get with (6.1.34) that, for large,
|Gk, (y ) |2
( n2 1)(12)
1
L H (y ) k,
Gk, (y )
(1 )
2
Gk, (y )

j1
( n 1)(12)
1
2 2
u (y )
i,2
Gi, (y )
i=1
( n2 1)(12)
+j,
Gl, (y )
p
( n2 1)(12)
j,
(1 )
Gi, (y )
|Gl, (y ) |2
2
Gl, (y )
2 2
u (y )
i=j
C12 ( n2 1)(12)
(2n)(1)2
dg (yk, , y )
C21 k,
j1

( n 1)(12)
(2n)(1)
2 2
2
1
u (y )
C2
i,
dg (yi, , y )
(1 )
i=1
C12 ( n2 1)(12)
(2n)(1)2
dg (y , yl, )
C21 j,
p

n
1
(12)
)
(
(2n)(1)
2 2
C21 j,2
dg (yi, , y )
u (y )
+ (1 )
i=j
C12
2
(1 ) 1
k, (y ) dg (yk, , y )
C2

j1

2 2
C21
i, (y ) u (y )
i=1
!
C12
( n 1)(12)
(2n)(1)2
(1 ) 1
+j,2
Rj, (y )
C2
"

2
2 2
.
C21 (p j + 1) Rj, (y ) u (y )
Since y k, , i, (y ) k, (y ) for all i {1, . . . , j 1}. Therefore,

2
L H (y ) k, (y ) dg (yk, , y )
( n2 1)(12)
+j,
A,
(2n)(1)2
Rj, (y )
B,
where
A, = (1 )
C12
2
2 2
1
(j 1) dg (yk, , y ) u (y )
1 C2
C2
and
B, = (1 )
C12
2
2 2
1
(p j + 1) Rj, (y ) u (y )
.
1 C2
C2
101
Since y k, and using (6.1.35), we have that

2
2 2
Rj, (y ) u (y )
2 2
dg (yk, , y ) u (y )
D () .
In particular, from the definition (6.1.22) of D (), L H (y ) 0 for large.

D ()
,
2
k, , and that dg (yk, , y ) Rj, (y ). Since y
k, , here again
that y
2 2
u (y )
dg (yk, , y )
2
2 2
D () dg (yk, , y ) u (y )
D ()
2
so that dg (yk, , y ) . From (6.1.2), (6.1.3), and (6.1.34),
|Gk, (y ) |2
( n2 1)(12)
1
Gk, (y )
(1 )
L H (y ) k,
2
Gk, (y )
j1

( n 1)(12)
1
2 2
i,2
Gi, (y )
u (y )
i=1
p
( n2 1)(12)
j,
Gi, (y )
2 2
u (y )
i=j
C12
2
(1 ) 1
k, (y ) dg (yk, , y )
C2

j1

2 2
C21
i, (y ) u (y )
i=1
( n 1)(12)
(2n)(1)
2 2
(p j + 1) C21 j,2
Rj, (y )
u (y )
for large. Since y k, , i, (y ) k, (y ) for all i {1, . . . , j 1}.
Thus,
2
L H (y ) k, (y ) dg (yk, , y )
(
(p j + 1) C21 j,
n
2 1
)(12)
A,
(2n)(1)
Rj, (y )
(6.1.36)
2 2
u (y )
where A, is as in case 2. Since Rj, (y ) dg (y , yk, ), and using the definition (6.1.32) of j, ,
( n 1)(12)
(2n)(1)
Rj, (y )
j,2
( n 1)(12)
(2n)(1)
j,2
dg (yk, , y )
(j, max {Aj, ; 1})( 2
1)(12)
(2n)(1)
dg (yk, , y )
By (6.1.28), Aj, j, j1, (1 + o (1)). Since j 1 j, we also have by

(6.0.7) that j, j1, . Hence,
j, max {Aj, ; 1} j1, (1 + o (1))
k, (1 + o (1))
102
CHAPTER 6
since k j 1. It follows that

( n 1)(12)
(2n)(1)
Rj, (y )
k, (y ) (1 + o (1)) .
j,2
2
L H (y ) k, (y ) dg (yk, , y )
C,
where
C, = (1 )
C12
2
2 2
1
(1 + o (1)) dg (yk, , y ) u (y )
.
1 pC2
C2
Since
2
2 2
dg (yk, , y ) u (y )
D () ,
and using the definition (6.1.22) of D (), we have that C, 0 for large.
Hence, L H (y ) 0 for large.
2 2
u (y )
D ()
2
and that y k, . Since y k, ,

2
2 2
dg (yk, , y ) u (y )
> D () .
Then (6.1.23) implies that

2
2 2
Rj, (y ) u (y )
D () .
It follows that Rj, (y ) and that Rj, (y ) dg (yk, , y ). We let l with

l {j, . . . , p} be such that Rj, (y ) = dg (yl, , y ). From (6.1.34),
j1

( n 1)(12)
1
2 2
2
L H (y )
i,
Gi, (y )
u (y )
i=1
2
( n 1)(12)
1 |Gl, (y ) |
+ (1 ) j,2
Gl, (y )
2
Gl, (y )
p
( n 1)(12)
1
2 2
j,2
Gi, (y ) u (y )
.
i=j
We can then use (6.1.2) and (6.1.3) to write that

j1

2 2
1
i, (y ) u (y )
L H (y ) C2
i=1
C12 ( n2 1)(12)
(2n)(1)2
dg (y , yl, )
1 j,
C2
p

n
( 1)(12)
(2n)(1)
2 2
C21 j,2
dg (yi, , y )
u (y )
+ (1 )
i=j
103
and then that

L H (y ) C21
j1

2 2
i, (y ) u (y )
i=1
( n2 1)(12)
+j,
(2n)(1)2
Rj, (y )
B,
where
B, = (1 )
C12
2
2 2
1
(p j + 1) Rj, (y ) u (y )
1 C2
C2
is as in case 2. Since y k, , i, (y ) k, (y ) for all i {1, . . . , j 1}.

Moreover, since y k, , it follows from the definition of Aj, that
k, (y ) (Aj, j, )( 2
1)(12)
(2n)(1)
Rj, (y )
By the definition (6.1.32) of j, , this leads to

( n 1)(12)
(2n)(1)
Rj, (y )
.
k, (y ) j,2
Coming back to the above estimate on L H (y ), we get that
( n 1)(12)
(2n)(1)2
Rj, (y )
L H (y ) j,2
#
$
C12
2
2 2
1
(1 ) 1 pC2 Rj, (y ) u (y )
.
C2
Since Rj, (y ) dg (yk, , y ), (6.1.23) and the definition (6.1.22) of D() give
that L H (y ) 0 for large.
It clearly follows from the above four cases that, for sufficiently large,
L H 0 in M \ pi=1 Byi, (R () i, ) .
(6.1.37)
Combining (6.1.37) with (6.1.33) and the fact that L u = 0 in M , the maximum
principle gives the existence of some C () > 0, independent of , such that for any
y M \ pi=1 Byi, (R () i, ), u (y) C () H (y). From (6.1.3) it follows
that there exists C () > 0, independent of , such that

j1

( n2 1)(12)
(2n)(1)
(6.1.38)
i, (y) + j,
Rj, (y )
u (y) C ()
i=1
for all y M \
pi=1
Byi, (R () i, ).
In order to end the proof of claim 6.1.3 (see the definition (6.1.32) of j, ) it
remains to prove that Aj, is bounded from above. We proceed by contradiction
and assume on the contrary that
Aj, +
(6.1.39)
as +. By the definition (6.1.26) of Aj, , there exist k {1, . . . , j 1}

and y k, such that
( n2 1)(12)
=
Aj,
k, (y )
( n 1)(12)
(2n)(1)
j,2
Rj, (y )
104
CHAPTER 6
Since Aj, + as +, we have that j, = Aj, j, . Thus

( n 1)(12)
(2n)(1)
k, (y ) = j,2
Rj, (y )
.
(6.1.40)
Combining (6.1.38) and (6.1.40) we then get that

j1

u (y ) C ()
i, (y ) + k, (y ) .
i=1
Since y
k, ,
i,
(y )
k,
(y ) for all i {1, . . . , j 1}. Hence,
u (y ) C () jk, (y ) .
Independently, since y k, ,
n
dg (yk, , y ) 2
u (y ) D ()
n2
4
It follows that
n2
n
D () 4
1
dg (yk, , y ) 2 k, (y )
C () j
( n2 1)(12)

k,
dg (yk, , y )
so that
dg (yk, , y ) = O (k, ) .
(6.1.41)

n 1 (12)
j, ( 2 )
(n2)(1)
(n2)(1)
=
dg (yk, , y )
Rj, (y )
k,

n
( 2 1)(12) n2 1
.
= O j,
k,
Since j, = Aj, j, = O (j1, ) by (6.1.28), and from (6.0.7), it follows that
Rj, (y ) = O (k, ) .
(6.1.42)
Let l {j, . . . , p} be such that Rj, (y ) = dg (yl, , y ). By (6.1.41) and (6.1.42)

we can write that
dg (yl, , yk, ) dg (yk, , y ) + dg (yl, , y )
= O (k, ) ,
a contradiction with (6.0.8). Therefore, (6.1.39) is false, and j, = O (j, ).
Then (6.1.38) reads as
j1

( n2 1)(12)
(2n)(1)
u (y) C ()
i, (y) + j,
Rj, (y )
i=1
and this ends the proof of claim 6.1.3.

The last claim of subsection 6.1.1 is as follows.
105
C LAIM 6.1.4. Assume that u0 0. If (6.1.19)j1 holds for some j = 2, . . . , p,

then (6.1.19)j holds also.
Proof of claim 6.1.4. Let (y ) be a sequence in M \ pi=1 Byi, (R0 i, ), where
R0 is as in (6.0.9). We have to prove that
lim sup j1
+
u (y )
0
i=1 i,
2n
2
(y ) + j,
Rj, (y )
< +
(6.1.43)
(6.1.44)
where
n
2n
2
0i, (y) = i,
dg (yi, , y)
2
small. Given i = 1, . . . , j 1, we let
We fix 0 < < n+2

( n2 1)(12)
(2n)(1)
Rj, (y)
i, = y i, , i, (y) j,
(6.1.45)
where i, is as (6.1.21), and i, is as in (6.1.18). We also let

i, .
j, = M \ j1
i=1
Let k {1, . . . , j 1}, and R () be as in claim 6.1.3. We claim that
By (R () k, )
k,
k,
(6.1.46)
(6.1.47)
and that for any i = k,

=
Byi, (R () i, )
k,
(6.1.48)
provided that is large. First we prove (6.1.47). Let (z ) be a sequence of points

in Byk, (R () k, ). We need to prove that k, (z ) i, (z ) for all i such
that i {1, . . . , j 1}, and that
( n 1)(12)
(2n)(1)
Rj, (z )
.
k, (z ) j,2
Given 1 i = k j 1, we assume by contradiction that i, (z ) k, (z ).

Since z Byk, R () k, ,

n
( 2 1)(12) n2 1
(n2)(1)
= O i,
k,
dg (yi, , z )

(n2)(1)
.
= O max {i, , k, }
Thus dg (yi, , z ) = O (max {i, , k, }), and
dg (yk, , yi, ) dg (z , yk, ) + dg (yi, , z )
= O (k, ) + O (max {i, , k, })
= O (max {i, , k, }) .
Such an equation is in contradiction with (6.0.8). Hence i, (z ) < k, (z )
for all i {1, . . . , j 1}, i = k, and z k, . Now we assume by contradiction
that
( n 1)(12)
(2n)(1)
k, (z ) j,2
Rj, (z )
.
106
CHAPTER 6
Since z Byk, R () k, and j k,

(n2)(1)
Rj, (z )
n

( 2 1)(12) n2 1
= O j,
k,

(n2)(1)
.
= O k,
Hence, Rj, (z ) = O (k, ). Let l {j, . . . , p} be such that Rj, (z ) is given

by Rj, (z ) = dg (yl, , z ). Then
dg (yk, , yl, ) dg (yl, , z ) + dg (yk, , z )
= O (k, ) ,
a contradiction with (6.0.8). This proves (6.1.47). Now we prove (6.1.48). Let
i {1, . . . , p}, i = k, and assume by contradiction that there exists a sequence
. Then dg (z , yi, ) = O (i, ), and,
(z ) of points in Byi, (R () i, )
k,
from (6.0.8), R (z ) = dg (yi, , z ). It easily follows that
( n 1)(12)
(2n)(1)
k, (z ) i,2
dg (yi, , z )
and then that
(n2)(1)
dg (yk, , z )
( n2 1)(12)
n
2 1
= O k,
i,

(n2)(1)
.
= O max {k, , i, }
As above, this is in contradiction with (6.0.8), so that (6.1.48) is proved.

Now we claim that, for any k {j, . . . , p},
Byk, (R () j, )
j,
(6.1.49)
and that, for any k {1, . . . , j 1},

=
Byk, (R () k, )
j,
(6.1.50)
provided that is large. First we prove (6.1.49). Let k {j, . . . , p} and let
(z ) be a sequence of points in Byk, (R () j, ). Assume by contradiction that
. Then, from the definition (6.1.46) of
, we get that there exists i
z
j,
j,
with i {1, . . . , j 1} such that
( n 1)(12)
(2n)(1)
i, (z ) j,2
Rj, (z )
.
Since
Rj, (z ) dg (yk, , z ) = O (j, )
and since i j, we can write, from (6.0.7), that
n 1 (12)

i, ( 2 )
(n2)(1)
(n2)(1)
Rj, (z )
dg (yi, , z )
j,

n
( 1)(12) n2 1
= O i,2
j,

(n2)(1)
.
= O i,
107
Thus dg (yi, , z ) = O (i, ), and

dg (yi, , yk, ) dg (yi, , z ) + dg (z , yk, )
= O (i, ) + O (j, )
= O (i, ) .
Noting that this is in contradiction with (6.0.8), we get that (6.1.49) is true. Independently, it is easily seen that (6.1.50) is a direct consequence of (6.1.47).
By claim 6.1.3, (6.1.47), and (6.1.48), we get that, for any k {1, . . . , j 1}
\By (R () k, ),
and any y
k,
k,
u (y) Ck, (y)
(6.1.51)
where C > 0 is independent of . Independently, it follows from claim 6.1.3 and

\ p By (R () i, ),
(6.1.50) that, for any y
i,
j,
i=j
( n 1)(12)
(2n)(1)
Rj, (y)
u (y) Cj,2
(6.1.52)
where C > 0 is independent of .

Now we prove (6.1.43). Let us assume first that there exists k {1, . . . , j 1}
such that
dg (yk, , y )
R
k,
as + for some R R0 . Then, from (6.0.8), (6.0.9) and point (P1) of
Theorem 4.1,

n
2n
2 1
.
dg (yk, , y )
u (y ) = O k,
In particular, it easily follows that (6.1.43) is true. As a consequence, we can
assume that, for any k {1, . . . , j 1},
dg (yk, , y )
+
k,
(6.1.53)
as +. Let us assume in addition that Rj, (y ) = O (j, ). We claim that

R (y ) = Rj, (y ). In order to prove this claim, we let l {j, . . . , p} be such
that Rj, (y ) = dg (y , yl, ). Then we can write that, for any k {1, . . . , j 1},
dg (yk, , y ) dg (yk, , yl, ) dg (yl, , y )
j,
j,
j,
dg (yk, , yl, ) k,
O (1) .
k,
j,
In particular, using (6.0.7) and (6.0.8), we get that
dg (yk, , y )
+
j,
108
CHAPTER 6
as +. Hence, dg (yk, , y ) > dg (yl, , y ) for large, and, as an immediate consequence, we get that R (y ) = Rj, (y ). The above claim is proved. By
(6.0.11) we can then write that

n 1
n
n
Rj, (y ) 2
n2 1 2
1
j, u (y ) =
R (y ) 2 u (y )
Rj, (y )
j,
= O (1) ,
and we have proved that, if Rj, (y ) = O (j, ), then (6.1.43) holds. Conversely,
we assume, in addition to (6.1.53), that
Rj, (y )
+
(6.1.54)
j,
as +. Let G be the Greens function of g + h0 , where h0 is as above;
namely, h0 C 0, (M ) is such that g + h0 is coercive and h > h0 . From the
Greens representation formula and equation (E ),
2 1
u (y )
G (y , y) u (y)
dvg
(6.1.55)
M
for large. Let k {1, . . . , j 1}, and

1
1
Bk, = y k, , dg (y , y) dg (yk, , y ) ,
2

1
2
Bk, = y k, , dg (y , y) < dg (yk, , y ) .

2
1
2
= B B . Independently, (6.1.47) and (6.1.53) give that

Clearly,
k,
k,
k,
1
Byk, (R () k, ) Bk,
for large. It follows that
2 1
G (y , y) u (y)
dvg
k,
2 1
Byk, (R()k, )
G (y , y) u (y)
2 1
1 \B
Bk,
yk, (R()k, )
2
Bk,
dvg
G (y , y) u (y)
2 1
G (y , y) u (y)
(6.1.56)
dvg
dvg .
1
,
From (6.1.3), for any y Bk,
2n
G (y , y) C2 dg (y , y)
Hence,
2n
C2 2n2 dg (yk, , y )
2 1
Byk, (R()k, )
G (y , y) u (y)
2n
Cdg (yk, , y )
2n
Cdg (yk, , y )
dvg
Byk, (R()k, )
2 1
u (y)
dvg
1

Volg Byk, (R () k, ) 2 u 22 1
(6.1.57)
109
and, since u 2 , we get that

2 1
G (y , y) u (y)
dvg = O 0k, (y ) .
(6.1.58)
Byk, (R()k, )
Using (6.1.51) and (6.1.57), we can also write that
2 1
G (y , y) u (y)
dvg
1 \B
Bk,
yk, (R()k, )
2n
Cdg (yk, , y )
Since <
2
n+2 ,
2 1
1 \B
Bk,
yk, (R()k, )
k, (y)
dvg .
it follows that
2 1
1 \B
Bk,
yk, (R()k, )
G (y , y) u (y)
2n
Cdg (yk, , y )
( n2 +1)(12)
k,
(n+2)(1)
M \Byk, (R()k, )
2n
Cdg (yk, , y )
dg (yk, , y )
dvg
( n2 +1)(12)(n+2)(1)+n
k,
.
In particular, here again,
2 1
1 \B
Bk,
yk, (R()k, )
dvg
G (y , y) u (y)

dvg = O 0k, (y ) .
(6.1.59)
2
,
Given y Bk,
dg (yk, , y) dg (yk, , y ) dg (y, y )

1
dg (yk, , y ) .
2
2
By (6.1.51) and (6.1.53) we then get that for any y Bk,
,
2 1
u (y)
( n2 +1)(12)
(n+2)(1)
Ck,
dg (yk, , y )
.
Using (6.1.3) we can write that
2 1
G (y , y) u (y)
dvg
2
Bk,
( n2 +1)(12)
(n+2)(1)
Ck,
dg (yk, , y )
( n2 +1)(12)
(n+2)(1)
Ck,
dg (yk, , y )
2n
dvg .
dg (y , y)
By ( 12 dg (yk, ,y ))
2
Bk,
2n
dg (y , y)
dvg
110
CHAPTER 6
Hence,
2
Bk,
2 1
G (y , y) u (y)
( n2 +1)(12)
2(n+2)(1)
dvg Ck,
dg (yk, , y )
C0k,

(y )
k,
dg (yk, , y )
2(n+2)
2
so that by (6.1.53), and since < n+2
,

2 1
G (y , y) u (y)
dvg = o 0k, (y ) .
(6.1.60)
Combining (6.1.58)(6.1.60) with (6.1.56), we proved that

2 1
G (y , y) u (y)
dvg = O 0k, (y )
(6.1.61)
2
Bk,
k,
for all k {1, . . . , j 1}. Given k {j, . . . , p}, we now let

\ p By (R () j, ) , Rj, (y) = dg (yk, , y)
k, = y
j,
i,
i=j
and define

1
= y k, , dg (y , y) Rj, (y ) ,
2

1
2
Bk, = y k, , dg (y , y) < Rj, (y ) .
2
1
Bk,
By (6.1.49),
.
pi=j Byi, (R () j, )
j,
Independently, it is clear that

p
j, = p By (R () j, )
k,
i,
i=j
k=j
and we can write
that
j,
2 1
G (y , y) u (y)
dvg
2 1
p
i=j Byi, (R()j, )
+
+
1
k=j Bk,
p
k=j
2
Bk,
G (y , y) u (y)
2 1
G (y , y) u (y)
2 1
G (y , y) u (y)
dvg
(6.1.62)
dvg
dvg .
1
Let i {j, . . . , p}. By (6.1.54), dg (yi, , y ) j,
+ as +. It follows
1
that if z Byi, (R () j, ), then dg (z , y ) dg (yi, , y ) 1 as +.
From (6.1.3) we can then write that, for large,
2n
G (y , y) C2 dg (y , y)
2n
C2 2n2 Rj, (y )
111
in
pi=j Byi,
(R () j, ). As a consequence,
2 1
G (y , y) u (y)
dvg
p
i=j Byi, (R()j, )
2n
CRj, (y )
2 1
p
i=j Byi, (R()j, )
and, from Holders inequality,
2 1
G (y , y) u (y)
p
i=j Byi, (R()j, )
2n
CRj, (y )
u (y)
dvg
dvg
1

u 22 1 Volg pi=j Byi, (R () j, ) 2
where C > 0 is independent of . Since u 2 , we get that
2 1
G (y , y) u (y)
dvg
p
i=j Byi, (R()j, )
n
2 1
2n
= O j, Rj, (y )
(6.1.63)

.
1
, from (6.1.3),
In Bk,
2n
G (y , y) C2 dg (y , y)
2n
C2 2n2 Rj, (y )
Independently, using (6.1.52) and the definition of k, ,

2 1
u (y)
( n +1)(12)
(n+2)(1)
Cj,2
dg (yk, , y)
.
2
,
As a consequence, since < n+2
2 1
G (y , y) u (y)
dvg
1
Bk,
2n
CRj, (y )
2n
CRj, (y )
2n
CRj, (y )
( n +1)(12)
j,2
( n +1)(12)
j,2
1
Bk,
(n+2)(1)
dg (yk, , y)
M \Byk, (R()j, )
dvg
(n+2)(1)
dg (yk, , y)
dvg
( n +1)(12)(n+2)(1)+n
j,2
.
It follows that

n
2 1
2n
2 1
.
G (y , y) u (y)
dvg = O j,
Rj, (y )
1
Bk,
2
In Bk,
, we have that
dg (yk, , y) dg (yk, , y ) dg (y , y)
1
Rj, (y ) Rj, (y )
2
1
Rj, (y )
2
(6.1.64)
112
CHAPTER 6
so that, by (6.1.52) and the definition of k, ,

2 1
u (y)
( n +1)(12)
(n+2)(1)
Cj,2
Rj, (y )
.
Using (6.1.3) again, we can write that
( n +1)(12)
2 1
(n+2)(1)
G (y , y) u (y)
dvg Cj,2
Rj, (y )
2
Bk,
2
Bk,
G (y , y) dvg
( n +1)(12)
(n+2)(1)
Cj,2
Rj, (y )
2n
dg (y , y)
dvg
2
Bk,
( n +1)(12)
(n+2)(1)
Cj,2
Rj, (y )
2n
dvg
dg (y , y)
By ( 12 Rj, (y ))
( n +1)(12)
2(n+2)(1)
Cj,2
Rj, (y )
2(n+2)

n
j,
2n
2 1
Cj, Rj, (y )
.
Rj, (y )
2
, we then get that
By (6.1.54), and since < n+2

n

2 1
2n
2 1
.
G (y , y) u (y)
dvg = o j,
Rj, (y )
(6.1.65)
Substituting (6.1.63)(6.1.65) into (6.1.62), it follows that
n

2 1
2n
2 1
Rj, (y )
G (y , y) u (y)
dvg = O j,
.
(6.1.66)
2
Bk,
j,
= M , we get with (6.1.55), (6.1.61), and (6.1.66) that

Noting that ji=1
i,
(6.1.43) is true. This proves claim 6.1.4.
2
By claim 6.1.2, (6.1.19)j is true for j = 1. By induction, thanks to claim 6.1.4,
we then get that (6.1.19)j is true for all j = 1, . . . , p. In particular, (6.1.19)p is
true. Noting that equations (6.1.1) and (6.1.19)p are the same, (6.1.1) is proved.
6.1.2 The case u0 0. We assume in this subsection that u0 0. Let R0 be as
in (6.0.9). We want to prove that there exists C > 0 and a sequence ( ), 0
as +, such that for any y M \ pj=1 Byj, (R0 j, ),
u (y) u0 (y) C
p
2n
2
j,
dg (yj, , y)
(6.1.67)
j=1
where the yj, s and j, s are as in (6.0.6). As in subsection 6.1.1, we split the
proof of (6.1.67) into several steps that we refer to as claims. Here again we proceed by induction. As in subsection 6.1.1, we let h0 C 0, (M ) be such that
113
h0 < h and g + h0 is coercive. We let G be the Greens function of the operator g + h0 . Our first claim is the following:
C LAIM 6.1.5. Assume that u0 0, and let 0 < < 21 . If is sufficientl small,
then there exist R() > 0 and C () > 0 such that for any > 0 and any y in
M \ pj=1 Byj, (R () j, ),
n

( 2 1)(12)
(2n)(1)
(2n)
R (y)
+ R (y)
u (y) C () 1,
(6.0.6).
Proof of claim 6.1.5. We fix 0 < < 12 and x0 S, where S is as in Theorem 4.1.
We let

(1 ) C12
(1 ) C12
;
(6.1.68)
D () = min
2(1)
4p
4p
C22
C
2
where C1 and C2 are given by (6.1.2) and (6.1.3). For as in (6.1.2), we let
0 < () < 2 and R () R0 be such that, for sufficiently large, and any
y Bx0 ( ()) \ pj=1 Byj, (R () j, ),

2
2 2
R (y) u (y)
+ h0 (y) h (y) D () and

(6.1.69)
2
2 2
+ (1 ) h0 (y) h (y) D () .
R (y) u (y)
The existence of () and R () easily follows from (6.0.12). Taking () sufficiently small, we may also assume that
dg (x0 , S\ {x0 }) 2 () .
(6.1.70)
Let L be the linear operator given by

2 2
L (u) = g u + h u u
u.
From [9], L satisfies the maximum principle. We let

p
p

1
1
2
H (y) =
Gj, (y)
and H (y) =
Gj, (y)
j=1
j=1
where Gj, (y) = G (yj, , y). Straightforward computations give that

p

1
2 2
1
L H (y) =
Gj, (y)
h (y) (1 ) h0 (y) u (y)
j=1
+ (1 )
|Gj, (y)|2
(y) =
p
(6.1.71)
Gj, (y)
and
L H2
Gj, (y)
2 2
h (y) h0 (y) u (y)
j=1
+ (1 )
|Gj, (y)|2
2
Gj, (y)
(6.1.72)
114
CHAPTER 6
in M \ {y1, , . . . , yp, }. We claim that, for sufficiently large,

L H1 0 and L H2 0
in Bx0 ( ()) \ pj=1 Byj, (R () j, ). In order to prove this claim, we let (y ) be
a sequence in Bx0 ( ()) \ pj=1 Byj, (R () j, ), and let k {1, . . . , p} be such
that dg (yk, , y ) = R (y ). It is clear that R (y ) 2 () < for large so
that, by (6.1.2),
|Gk, (y)|2
Gk, (y)
C12
C12
dg (y , yk, )
R (y )
Then, by (6.1.69),
C12
L H1 (y ) (1 ) Gk, (y )
2
D () R (y )
R (y )
p
Gj, (y )
j=1
and
L H2 (y ) (1 ) Gk, (y )
C12
D () R (y )
R (y )
p
Gj, (y )
j=1
for large. From (6.1.3),

1
2n
dg (yk, , y )
C2
1
2n
R (y )
C2
Gk, (y )
and, for any j {1, . . . , p},

2n
Gj, (y ) C2 dg (yj, , y )
2n
C2 R (y )
Hence,
L H1
(2n)(1)2
C12
pD () C21
(1 ) 1
C2
(y ) R (y )
and
(2n)2
L H2 (y ) R (y )

(1 )
C12
pD () C2
C2

.
Coming back to the definition (6.1.68) of D (), it follows that L H1 (y ) 0

and L H2 (y ) 0 for large enough. Summarizing, we proved that
L H1 0 , L H2 0 in Bx0 ( ()) \ pj=1 Byj, (R () j, )
(6.1.73)
115
for large. It easily follows from point (P1) of Theorem 4.1, (6.0.7), (6.0.9), and
(6.1.3), that there exists C () > 0, independent of , such that for any j with
j {1, . . . , p},
( n2 1)(12)
1
Gj, (y)
on Byj, (R () j, ) .
u (y) C () 1,
(6.1.74)
Independently, it easily follows from (6.1.70) and point (P1) of Theorem 4.1 that
u (y) C () H2 (y) on Bx0 ( ())
(6.1.75)
for some C () > 0 independent of . Combining (6.1.73)(6.1.75), we get with

the maximum principle that
# n
$
( 2 1)(12) 1
u (y) C () 1,
H (y) + H2 (y)
in Bx0 ( ()) \ pj=1 Byj, (R () j, ). Since x0 S is arbitrary, and considering
(6.1.3), this proves that the estimate of claim 6.1.5 holds when we replace M by a
()-neighborhood of S. As a straightforward consequence of point (P1) of Theorem 4.1, we also have that such an estimate holds outside this neighborhood of S.
This ends the proof of claim 6.1.5.
2
Our next claim is the following:
C LAIM 6.1.6. Assume that u0 0. There exist C > 0 and a sequence
( ), 0 as +, such that for any > 0 and any y in M \ pj=1
Byj, (R0 j, ),
n
2n
2
u (y) u0 (y) C1,
R (y)

(6.0.6).
Proof of claim 6.1.6. For j = 1, . . . , p, we let yj = xNj be the point of Rn given
by point (P1) of Theorem 4.1. We let
D0 = 2 sup
sup |z|n2 u (z yj )
j=1,...,p zRn
(6.1.76)
where u is given by (4.1.5). It easily follows from point (P1) of Theorem 4.1 that,
up to a subsequence,
sup
M \p
j=1 Byj, ( )
|u u0 | 0 as +
(6.1.77)
where ( ) is a sequence of positive real numbers such that 0 as +.

We claim that there exists C > 0 such that for any sequence (y ) of points in
M \ pj=1 Byj, (R0 j, ),

n
2n
2 1
lim sup u (y ) u0 (y ) C1,
R (y )
0.
(6.1.78)
+
In order to prove (6.1.78), we distinguish four cases.

Case 1. We assume that R (y ) as + for some > 0. Then
dg (y , S) as + and point (P1) of Theorem 4.1 gives that (6.1.78)
holds with C > 0.
116
CHAPTER 6
Case 2. We assume that there exists k {1, . . . , p} such that for some R R0 ,
dg (yk, , y )
R
k,
as +. By (6.0.9) and point (P1) of Theorem 4.1, we have in that case that
n
1
lim 2 u
+ k,
(y ) = u (y0 yk )
where, up to a subsequence,
y0 = lim
Independently,
n
2
lim k,
1
exp1
yk, (y ) .
k,
2n
2
k,
dg (yk, , y )
= |y0 |2n
so that
lim
u (y )
n
2 1
2n
k, dg (yk, , y )
= |y0 |n2 u (y0 yk ) .

u (y )
lim
n
2 1
2n
1, R (y )
1
D0
2
where D0 is as in (6.1.76). Since

R (y ) dg (yk, , y ) 2Rk, ,
we get with (6.0.7) that
n
2n
2
R (y )
1,
as +. It follows that (6.1.78) holds with C D0 .

Case 3. We assume that, for any j {1, . . . , p},
dg (yj, , y )
+
j,
as +. We also assume that R (y ) = O (1, ). Using (6.0.12) we then
get that
1 n
|u (y ) u0 (y ) |
1, 2 R (y )

n2 1
n
R (y )
1
=
R (y ) 2 |u (y ) u0 (y ) |
1,

n
1
= O R (y ) 2 |u (y ) u0 (y ) |
n2
= o (1) .
It follows that (6.1.78) holds with C > 0.
117
R (y )
+
1,
as +. We let G be the Greens function of g + h . We write with the
Greens representation formula that

2 1
2 1
0
u (y ) u (y ) =
G (y , y) u (y)
u0 (y)
dvg
M
(6.1.79)
+
G (y , y) (h (y) h (y)) u0 (y) dvg .
R (y ) 0 and
We let C2 > 0 (we refer to Appendix A) be such that, for any > 0 and any
x, y M , x = y,
1
2n
2n
dg (x, y)
G (x, y) C2 dg (x, y)
.
(6.1.80)
C2
We clearly have that
G (y , y) (h (y) h (y)) u0 (y) dvg = o (1)

(6.1.81)
M
and that
M \p
i=1 Byi, ( )

2 1
2 1
dvg = o (1)
G (y , y) u (y)
u0 (y)
where is as in (6.1.77). Coming back to (6.1.79), using (6.1.81) and the above
equation, we get that
2 1
0
u (y ) u (y )
G (y , y) u (y)
dvg + o (1) . (6.1.82)
p
i=1 Byi, ( )
Given j = 1, . . . , p, we now let

j, = {y M s.t. R (y) = dg (yj, , y)}
and let
1
Bj,
2
Bj,

1
= y j, Byj, ( ) s.t. dg (y , y) R (y ) ,
2

1
= y j, Byj, ( ) s.t. dg (y , y) < R (y ) .
2
1
We fix 0 < < n+2
small. Let j {1, . . . , p}. Up to increasing so that
1, = o( ), we can write that
2 1
G (y , y) u (y)
dvg
j, Byj, ( )
2 1
j, Byj, (R()1, )
G (y , y) u (y)
1 \B
Bj,
yj, (R()1, )
2 \B
Bj,
yj, (R()1, )
dvg
2 1
G (y , y) u (y)
2 1
G (y , y) u (y)
(6.1.83)
dvg
dvg
118
CHAPTER 6
1
where R () is given by claim 6.1.5. Since R (y ) 1,
+ as +, we
can write that, for large and any y j, Byj, (R () 1, ),
dg (y , y) dg (y , yj, ) dg (yj, , y)
R (y ) R () 1,
1
R (y ) .
2
Since u 2 , using (6.1.80) and Holders inequality, we have that, for
large,
2 1
G (y , y) u (y)
dvg
j, Byj, (R()1, )
2n
C2 2n2 R (y )
2 1
j, Byj, (R()1, )
u (y)
dvg

1

u 22 1 Volg Byj, (R () 1, ) 2

21
2n 2 1 n1
n n
R () 1,
2n1 C2 R (y )
n

1

n
n
1
2n
n1
2 1 n1 2
2 1
1,
2
C2
R () 2
R (y )
.
n
2n
C2 2n2 R (y )
Thus, for large,
2 1
j, Byj, (R()1, )
G (y , y) u (y)
n
2 1
dvg
(6.1.84)
2n
D1 1, R (y )
where
D1 = 2n1 C2 2
n1
21
R () 2
(6.1.85)
1
Given y Bj,
, we get from (6.1.80) that
2n
G (y , y) C2 2n2 R (y )
1
and, if y Bj,
\Byj, (R () 1, ), we get from claim 6.1.5 that
2 1
u (y)
22
( n2 +1)(12)
(n+2)(1)
dg (yj, , y)
1,
"
(n+2)
+dg (yj, , y)
.
2 1
C ()
Using (6.0.8) we indeed have that

Byi, (R () i, ) j, =
119
when is large and i = j. Thus we can write that
2 1
G (y , y) u (y)
dvg
1 \B
Bj,
yj, (R()1, )
22
2 1
C ()
(n+2)
1 \B
Bj,
yj, (R()1, )
2 1
+22 +n4 C ()
2n
C2 R (y )
1
n+2 ,
dg (yj, , y)
dvg .
and since 0 as +, it is easily checked with

(n+2)
1 \B
Bj,
yj, (R()1, )
G (y , y) dg (yj, , y)
dvg
( n2 +1)(12)
1,
(n+2)(1)
1 \B
Bj,
yj, (R()1, )
Noting that <

(6.1.80) that
G (y , y) dg (yj, , y)
G (y , y)
n
n1
n1
n
dvg

= O 1(n+2)
= o (1) .
Similarly, it is easily checked that
1 \B
Bj,
yj, (R()1, )
n1
n(n+2)
Byj, ( )
dvg
dg (yj, , y)
(n+2)(1)
dg (yj, , y)
dvg
dvg
2n1
n(n+2)(1)
(R () 1, )
(n + 2) (1 ) n
for large. Therefore,
2 1
G (y , y) u (y)
dvg
1 \B
Bj,
yj, (R()1, )
n
2 1
(6.1.86)
2n
D2 1, R (y )
where

D2 = 22 +n3
+ o (1)
n1
2 1
n(n+2)(1)
C ()
C2 R ()
. (6.1.87)
(n + 2) (1 ) n
2
, we have that
In Bj,
dg (yj, , y) dg (yj, , y ) dg (y, y )
1
R (y ) R (y )
2
1
R (y )
2
so that, applying claim 6.1.5, we get that
!

2 1
2 1
(n+2)
dg (yj, , y)
22 2 C ()
u (y)
"
( n2 +1)(12)
(n+2)(1)
.
+2(n+2)(1) 1,
R (y )
120
CHAPTER 6
Once again, we used the fact that Byi, (R () i, ) j, = for large and any
i = j. Using (6.1.80) we then get that
2 1
G (y , y) u (y)
dvg
2 \B
Bj,
yj, (R()1, )
22
2+(n+2)(1)
2 \B
Bj,
yj, (R()1, )
2 2
+2
22
2 1
C ()
2 1
2 1
C ()
By (
2 1
dg (y , y)
)

G (y , y) n1 dvg
M
= O 1,

n(n+2)
dvg
2(n+2)(1)
R (y )
2n
2
R (y )
= O 1,
n1
n
n1
dg (yj, , y)
( n2 +1)(12)
dvg
dvg
n
C ()
Byj, ( )
G (y , y) dg (yj, , y)
( n2 +1)(12)
(n+2)(1)
1,
R (y )
2n
R (y )
2
(n+2)
2 \B
Bj,
yj, (R()1, )
+22
G (y , y) dvg
C ()
2+(n+2)(1)
C2
( n2 +1)(12)
(n+2)(1)
1,
R (y )
1,
R (y )

+ o (1)

2(n+2)
+ o (1) .
1
+ as +, it follows that
Since R (y ) 1,
2 1
G (y , y) u (y)
dvg
2 \B
Bj,
yj, (R()1, )
n
2 1
2n
= o 1, R (y )
(6.1.88)
+ o (1) .
Substituting (6.1.84), (6.1.86), and (6.1.88) into (6.1.83), we then get that, for any
j {1, . . . , p},
2 1
G (y , y) u (y)
dvg
j, Byj, ( )
(6.1.89)
n
2n
2
(D1 + D2 + o (1)) 1,
R (y )
+ o (1)
where D1 and D2 are given by (6.1.85) and (6.1.87). Since M = pj=1 j, and
Byi, ( ) j, Byj, ( ) j, ,
n
2n
2
u (y ) u0 (y ) p (D1 + D2 + o (1)) 1,
R (y )
+ o (1) .
121
It follows that (6.1.78) holds with C > p (D1 + D2 ), where D1 is as in (6.1.85),

and D2 is as in (6.1.87).
From cases 14, for any sequence (y ) in M \ pj=1 Byj, (R0 j, ),

n
2n
2 1
0
R (y )
lim sup u (y ) u0 (y ) C1,
+
if C > 0 is chosen such that C D0 + p (D1 + D2 ), where D0 is as in (6.1.76),

D1 is as in (6.1.85), and D2 is as in (6.1.87). Clearly, this ends the proof of claim
6.1.6.
2
For j {1, . . . , p}, we say that (6.1.90)j holds if there exist C > 0, independent
of , and a sequence ( ), 0 as +, such that for any > 0 and any
y M \ pi=1 Byi, (R0 i, ),
u (y) u0 (y)
j1

n 1
n
1
2n
2n
2
2
C
i,
dg (yi, , y)
+ j,
Rj, (y)
+
(6.1.90)j
i=1
where Rj, is defined by (6.1.18). By claim 6.1.6, (6.1.90)j holds for j = 1. The
subject of claims 6.1.7 and 6.1.8 below is to prove that, for any j {2, . . . , p},
(6.1.90)j1 holds (6.1.90)j holds.
(6.1.91)
First we set up notation. We let j {2, . . . , p} and we assume that (6.1.90)j1

1
holds. We fix 0 < < n+2
such that h > (1 ) h0 . For any x S, S as in
Theorem 4.1, we let
Aj (x) = {i {1, . . . , j 1} s.t. yi, x as +} ,
Bj (x) = {i {j, . . . , p} s.t. yi, x as +} .
(6.1.92)
For k = 1, . . . , j 1, we let

(6.1.93)
k, = y M s.t. k, (y) i, (y) for all i {1, . . . , j 1}
where
( n 1)(12)
(2n)(1)
dg (yi, , y)
.
i, (y) = i,2
We also let

D () = min
(1 ) C12
(1 ) C12
;
2(1)
4p
4p
C22
C
(6.1.94)
(6.1.95)
where C1 and C2 are given by (6.1.2) and (6.1.3).

Assuming (6.1.90)j1 , we claim now that there exist R () R0 and () > 0
such that the following assertions hold:
(A1) for any k {1, . . . , j 1}, and any y M \ pi=1 Byi, (R () i, ), if
y k, , then
1

n
D () 2 2
0
2 1
(6.1.96)
u (y) u (y) <
min {Rj, (y) , dg (yk, , y)}
100
122
CHAPTER 6
when is sufficiently large;

(A2) for any x S,
2
2 2
() u0 (y)
D ()
100
(6.1.97)
when y Bx ( ());
(A3) for any x S, and any y Bx ( ()) \ pi=1 Byi, (R () i, ),
!
"
2
2 2
h (y) + h0 (y) D ()
(6.1.98)
R (y) u (y)
when is sufficiently large.
The proof of (A1) is based on (6.1.90)j1 and (6.0.12). Let (y ) be a sequence
of points in M \ pi=1 Byi, (Ri, ), R > R0 to be fixed later on, such that y
k, for some k {1, . . . , j 1}. In order to prove (A1) we have to prove that,
choosing R sufficiently large,

n
1
u (y ) u0 (y )
min {Rj, (y ) , dg (yk, , y )} 2
1

(6.1.99)
D () 2 2
<
100
for large. From (6.1.90)j1 ,
u (y ) u0 (y )
j2

n 1
n
2n
2n
2
2 1
C
i, dg (yi, , y )
+ j1, Rj1, (y )
+ o (1) .
i=1
Since
Rj1, (y ) = min {Rj, (y ) ; dg (yj1, , y )} ,
it is clear that
2n
Rj1, (y )
2n
Rj, (y )
2n
+ dg (yj1, , y )
Since y k, , dg (yi, , y ) Ri, , and

n
2n
2
i,
dg (yi, , y )
(n2)
= i, (y ) i,
(2n)
dg (yi, , y )
it follows that
u (y ) u0 (y )
j1

n 1
n
2n
2n
2
2 1
C
+ j1, Rj, (y )
+ o (1)
i, dg (yi, , y )

i=1
C R(2n)
j1

i=1

n
2 1
2n
i, (y ) + j1, Rj, (y )
+ o (1)
"
!
n
2n
2 1
Rj, (y )
+ o (1) .
C (j 1) R(2n) k, (y ) + j1,
123
Therefore,

u (y ) u0 (y )

n2 1

n
j1,
(2n)
2 1
dg (yk, , y )
k, (y ) +
+ o (1)
C (j 1) R
Rj, (y )

n2 1

n
j1,
+ o (1) .
C (j 1) R1 2 +
Rj, (y )
n
min {Rj, (y ) , dg (yk, , y )} 2

1
+ as +, then (6.1.99) holds with R sufficiently
If Rj, (y ) j1,
large. Without loss of generality, we may thus assume that
Rj, (y ) = O (j1, ) .
(6.1.100)
We claim that this implies that Rj, (y ) = R (y ). In order to prove this

claim, we let l {j, . . . , p} be such that Rj, (y ) = dg (y , yl, ). For any
i {1, . . . , j 1}, from (6.0.7) and (6.1.100),
dg (y , yi, ) dg (yl, , yi, ) dg (y , yl, )
j1,
j1,
j1,
dg (yl, , yi, ) i,
Rj, (y )
i,
j1,
j1,
dg (yl, , yi, )
O (1) .
i,
By (6.0.8), we then get that
dg (y , yi, )
+
j1,
as +, and it follows from (6.1.100) that, for any i {1, . . . , j 1}, we
have that dg (yi, , y ) > Rj, (y ) when is large. This proves that Rj, (y ) is
such that Rj, (y ) = R (y ) for large. Using (6.0.12), we then get that
n
Rj, (y ) 2
|u (y ) u0 (y ) | R + o (1)
where R 0 as R +. In particular, choosing R sufficiently large, (6.1.99)

holds. This proves (A1). An elementary remark is that one can choose () > 0
small such that (6.1.97) holds. This proves (A2). Another simple remark is that
(6.1.98) follows from (6.0.12). This proves (A3).
We now let Aj, be defined by
( n2 1)(12)
Aj,
=
sup
sup
i=1,...,j1 yi,
i, (y)
( n 1)(12)
(2n)(1)
j,2
Rj, (y)
(6.1.101)
where

i, = y i, \ pk=1 Byk, (R () k, ) s.t.
1

n
D () 2 2
0
2 1
.
u (y) u (y)
dg (yi, , y)
100
(6.1.102)
124
CHAPTER 6
By convention, Aj, = if the i, s, i = 1, . . . , j 1, are empty. We claim

that for R () sufficiently large,
j,
1.
(6.1.103)
lim sup Aj,
+
j1,
We prove (6.1.103) in what follows. We assume that Aj, = . Then, up to a
subsequence, there exist k {1, . . . , j 1} and y k, such that
k, (y ) = (Aj, j, )( 2
1)(12)
(2n)(1)
Rj, (y )
We proceed by contradiction. Assuming that (6.1.103) is false, we have that, up to

a subsequence,
( n2 1)(12)
(2n)(1)
Rj, (y )
.
k, (y ) j1,
(6.1.104)
From (6.1.90)j1 , and since y k, , this implies that

1

n
D () 2 2
1
dg (yk, , y ) 2
u (y ) u0 (y )
100
j1
n 1
n
2n
2
2 1
Cdg (yk, , y )
i,
dg (yi, , y )
i=1
n
2 1
2n
+j1, Rj, (y )
+ o (1)

Cdg (yk, , y )
n
2 1
R ()
(2n)
2n
i, (y )
i=1

2
Rj, (y )
+j1,
j1
+ o (1)

Cdg (yk, , y )
n
2 1
( n2 1)(12)
+j1,
(2n)
(j 1) R ()
(2n)(1)
Rj, (y )

Cdg (yk, , y )

+k,
(y )

CR ()
1 n
2
n
2 1
j1,
Rj, (y )
(2n)
(j 1) R ()
j1,
Rj, (y )
k, (y )
(n2)
+ o (1)
k, (y )
(n2)
(j 1) + R ()
+ o (1)
(n2)
j1,
Rj, (y )
(n2)
+ o (1) .
Taking R () sufficiently large, this implies in turn that

Rj, (y ) = O (j1, ) .
(6.1.105)
125
Coming back to (6.1.104), we get with (6.1.105) that

(n2)(1)
dg (yk, , y )
n 1 (12)
k, ( 2 )
(n2)(1)
Rj, (y )
j1,
n

( 2 1)(12) n2 1
= O k,
j1, .

Since k j 1, j1, k, . Thus,

dg (yk, , y ) = O (k, ) .
(6.1.106)
We let l {j, . . . , p} be such that Rj, (y ) = dg (yl, , y ). Using (6.1.105) and

(6.1.106) we can write that
dg (yk, , yl, ) dg (yk, , y ) + dg (yl, , y )
= O (k, ) + O (j1, )
= O (k, ) ,
a contradiction with (6.0.8). This proves (6.1.103). We now let
j, = j, max {Aj, ; 1} .
(6.1.107)
Claim 6.1.7 can then be stated as follows:

1
C LAIM 6.1.7. Assume that u0 0. We let 0 < < n+2
sufficientl small such
that h > (1 ) h0 , and assume that (6.1.90)j1 holds for some j {2, . . . , p}.
Let R () > R0 be such that (6.1.96)(6.1.98) and (6.1.103) hold. Then there
exists C () > 0 such that for any y M \ pi=1 Byi, (R () i, ),
u (y)
C ()
j1
i,
( n2 1)(12)
(y) + j,

(2n)(1)
Rj, (y)
(2n)
+ R (y)
i=1
where R is as in (6.0.10), Rj, is as in (6.1.18), the j, s are as in (6.0.6), the

yj, s are as in (6.0.6), i, is as in (6.1.94), and j, is as in (6.1.107).
Proof of claim 6.1.7. Let L be the linear operator given by
2 2
L (u) = g u + h u u
u,
and
H (y) =
j1

( n 1)(12)
1
i,2
Gi, (y)
i=1
( n 1)(12)
1
+j,2
Gi, (y)
p
i=j
p

i=1
Gi, (y)
126
CHAPTER 6
where Gi, (y) = G (yi, , y), and G is the Greens function of g + h0 . We fix
x0 S, where S is as in Theorem 4.1. We let
0 < () < ,
2
1
() < dg (x0 , S\ {x0 })
2
be such that (6.1.97) and (6.1.98) hold, where is as in (6.1.2), (6.1.3). Since
h > (1 ) h0 , direct computations give that, for any y M \ {y1, , . . . , yp, }
and for large,
j1

( n 1)(12)
1
L H (y)
i,2
Gi, (y)
i=1
|Gi, (y) |2
(1 )
Gi, (y)

2 2
u (y)
( n 1)(12)
1
Gi, (y)
+ j,2
p
i=j
|Gi, (y) |2
(1 )
+
p
Gi, (y)
Gi, (y)
(6.1.108)
2 2
u (y)
2 2
h (y) h0 (y) u (y)
i=1
+ (1 )

|Gi, (y) |2
.
2
Gi, (y)
We claim that, for sufficiently large,
L H 0 in Bx0 ( ()) \ pi=1 Byi, (R () i, ) .
Let (y ) be a sequence of points in Bx0 ( ()) \ pi=1 Byi, (R () i, ). We let
also l {1, . . . , p} be such that R (y ) = dg (yl, , y ). Since () < 2 , it is
clear that R (y ) for large. Thus we can apply (6.1.2) and (6.1.3). We get
that

p

|Gi, (y ) |2
2 2
Gi, (y ) h (y ) h0 (y ) u (y )
+ (1 )
2
Gi, (y )
i=1
1
C12
(2n)
dg (yl, , y )
(1 )
2
C2
dg (yl, , y )
p
"
!
2 2
Gi, (y ) u (y )
h (y ) + h0 (y ) .
i=1
From (6.1.98),
p
Gi, (y )
2 2
u (y )
"
h (y ) + h0 (y )
i=1
2
D () R (y )
p

i=1
Gi, (y )
127
so that, by (6.1.3),
p
Gi, (y )
2 2
u (y )
"
h (y ) + h0 (y )
i=1
2
D () R (y )
D () pC2 R
C2
p
(2n)
dg (y , yi, )
i=1
2+(2n)
(y )
Therefore, coming back to the definition (6.1.95) of D (),

p
2 2
Gi, (y ) h (y ) h0 (y ) u (y )
i=1
|Gi, (y ) |2
+ (1 )
(6.1.109)
Gi, (y )
3pD () C2 R
2+(2n)
(y )
for large. Now we estimate the other terms in the right-hand side of (6.1.108).
Up to a subsequence, we may assume that y k, for some k {1, . . . , j 1}.
We distinguish four cases.
Case 1. We assume that min {Rj, (y ) ; dg (yk, , y )} as +
for some > 0. Then, by (6.1.96), u (y ) = O (1). From (6.1.3), and since
y k, , this gives in turn that
j1

j1

( n2 1)(12)
2 2
1
u (y )
i,
Gi, (y )
=O
i, (y )
i=1
=O
i=1
k,

(y ) .
In particular, we can write that

2 2
u (y )
j1

( n 1)(12)
1
i,2
Gi, (y )
= o (1)
(6.1.110)
i=1
since dg (yk, , y )
that
for large, and k, 0 as +. Similarly, one has

( n 1)(12)
1
j,2
Gi, (y )
p
2 2
u (y )
i=j

n
( 1)(12)
(2n)(1)
= O j,2
Rj, (y )

n
( 1)(12)
.
= O j,2
By (6.1.103) [see also (6.1.107)] j, 0 as +. Hence,
( n 1)(12)
1
j,2
Gi, (y )
= o (1) .
p
2 2
u (y )
i=j
(6.1.111)
128
CHAPTER 6
Coming back to (6.1.108), we get with (6.1.109)(6.1.111) that for sufficiently

large, L H (y ) 0.
Case 2. We assume that y k, , that dg (yk, , y ) Rj, (y ), and that
Rj, (y ) 0 as +. With these assumptions, from (6.1.2), (6.1.3), and
(6.1.96), it is not difficult to check that, for large,
p
( n 1)(12)
1
j,2
Gi, (y )

i=j
(1 )
|Gi, (y ) |2
(6.1.112)
2 2
u (y )
Gi, (y )
( n 1)(12)
2+(2n)(1)
(3p + j 1) D () C21 j,2
Rj, (y )
.
Since y k, ,
1

n
D () 2 2
0
2 1
u (y ) u (y )
dg (yk, , y )
100
so that
2
2 2
dg (yk, , y ) u (y )

D ()

(6.1.113)
2 0
2 2
2 3
+ dg (yk, , y ) u (y )
.
max 1; 2
100
If k Aj (x0 ), Aj (x0 ) as in (6.1.92), then, since () < 12 dg (x0 , S\ {x0 }), we
get that dg (yk, , y ) as + for some > 0. In particular, we get with
(6.1.113) that u (y ) = O (1). As in the first case, we can then prove that
j1

( n 1)(12)
2 2
1
i,2
Gi, (y )
= o (1)
(6.1.114)
u (y )
i=1
and (6.1.114) holds under the assumption that k Aj (x0 ). Let us now assume
that k Aj (x0 ). Then dg (yk, , y ) 2 () for large so that, combining
(6.1.97) and (6.1.113),
2
2 2
D ()
dg (yk, , y ) u (y )
for large. From (6.1.2) and (6.1.3), this implies that

j1

|Gi, (y ) |2
( n2 1)(12)
1
2 2
i,
Gi, (y )
u (y )
(1 )
2
Gi, (y )
i=1
( n2 1)(12)
k,
dg
C21
(2n)(1)
(yk, , y )
2
C21 D () dg (yk, , y )
(1 )
i=1
dg (yk, , y )
j1

( n 1)(12)
(2n)(1)
i,2
dg (yi, , y )
C12
2
(1 ) 1
dg (yk, , y ) k, (y )
C2
C21 D () dg (yk, , y )
C12
j1

i=1
i, (y ) .
129
Since y k, , we then get that, when k Aj (x0 ),

j1

( n 1)(12)
1
i,2
Gi, (y )
i=1
(1 )
|Gi, (y ) |2
2
Gi, (y )

2 2
u (y )
(6.1.115)
(4p (j 1)) C21 D () dg (yk, , y ) k, (y ) .

Coming back to (6.1.108), it follows from (6.1.109), (6.1.112), (6.1.114), and
(6.1.115), that for sufficiently large L H (y ) 0.
Case 3. We assume that y k, , that dg (yk, , y ) Rj, (y ), and that
dg (yk, , y ) 0 as +. Using (6.1.2), (6.1.3), and (6.1.96), we can then
write that

j1

|Gi, (y ) |2
( n2 1)(12)
1
2 2
i,
Gi, (y )
u (y )
(1 )
2
Gi, (y )
i=1
(1 )
C12
2
dg (yk, , y ) k, (y )
C21
2
D () C21 dg (yk, , y )
j1
i, (y ) .
i=1
Since y k, , this implies that

j1

( n 1)(12)
1
i,2
Gi, (y )
i=1
(1 )
|Gi, (y ) |2
2
Gi, (y )

2 2
u (y )
(6.1.116)
(4p (j 1)) D () C21 dg (yk, , y ) k, (y ) .

Similarly, since Rj, (y ) dg (yk, , y ), we can write using (6.1.3) and (6.1.96)
that
p

( n 1)(12)
2 2
1
j,2
u (y )
Gi, (y )
i=j
( n 1)(12)
2
(2n)(1)
(p j + 1) D () C21 j,2
dg (yk, , y ) Rj, (y )
( n 1)(12)
2+(2n)(1)
(p j + 1) D () C21 j,2
dg (yk, , y )
.
By (6.0.7) and (6.1.103), since k j 1 j,
j, k, (1 + o (1)) .
Therefore,
p

( n 1)(12)
2 2
1
u (y )
Gi, (y )
j,2
(6.1.117)
i=j
2
(p j + 1) D () C21 dg (yk, , y )
k, (y ) (1 + o (1)) .
130
CHAPTER 6
Coming back to (6.1.108), we get with (6.1.109), (6.1.116), and (6.1.117) that,
for sufficiently large, L H (y ) 0.
Case 4. We assume that y k, and that min {Rj, (y ) ; dg (yk, , y )} 0
as +. Since y k, ,
n
dg (yk, , y ) 2

u (y ) u0 (y )
D ()
100
212
.
Using (6.1.96) we then get that

Rj, (y ) dg (yk, , y )
and, since Rj, (y ) 0, we have that
2
2 2
Rj, (y ) u (y )
D ()
+ o (1) .
100
From (6.1.2) and (6.1.3), this implies that

p
|Gi, (y ) |2
( n2 1)(12)
1
2 2
j,
(1 )
Gi, (y )
u (y )
2
Gi, (y )
i=j
( n 1)(12)
(2n)(1)2
(3p + j 1) C21 D () j,2
Rj, (y )
(6.1.118)
for large. On the other hand, still using (6.1.3), and since y
2 2
u (y )
k, ,
j1

( n 1)(12)
1
i,2
Gi, (y )
i=1
2
D () Rj, (y )
C21
j1
i, (y )
i=1
2 1
C2 k,
(j 1) D () Rj, (y )
(y ) .
We assumed that y belongs to k, . By the definition (6.1.101) of Aj, and the

definition (6.1.107) of j, , this implies that
k, (y ) (Aj, j, )( 2
1)(12)
(2n)(1)
Rj, (y )
( n 1)(12)
(2n)(1)
j,2
Rj, (y )
.
Hence, for large,
2 2
u (y )
j1

( n 1)(12)
1
i,2
Gi, (y )
(6.1.119)
i=1
(
(j 1) D () C21 j,
n
2 1
)(12)
2+(2n)(1)
Rj, (y )
Coming back to (6.1.108), we get with (6.1.109), (6.1.118), and (6.1.119) that,
for sufficiently large, L H (y ) 0.
131
It clearly follows from cases 14 that, for sufficiently large,

L H 0 in Bx0 ( ()) \ pi=1 Byi, (R () i, ) .
(6.1.120)
Independently, using point (P1) of Theorem 4.1 and (6.1.3), we easily get the existence of some C () > 0 such that

u C () H on Bx0 ( ()) \ pi=1 Byi, (R () i, ) .
(6.1.121)
Since L u = 0, noting that L satisfies the maximum principle (see [9]), it follows from (6.1.120) and (6.1.121) that
u C () H in Bx0 ( ()) \ pi=1 Byi, (R () i, ) .
From (6.1.3), this proves that the estimate of claim 6.1.7 holds when we replace M
by a ()-neighborhood of S. As a straightforward consequence of point (P1) of
Theorem 4.1, we also have that such an estimate holds outside this neighborhood
of S. This ends the proof of claim 6.1.7.
2
The last claim of subsection 6.1.2 is as follows.
C LAIM 6.1.8. Assume that u0 0. If (6.1.90)j1 holds for some j = 2, . . . , p;
then (6.1.90)j holds also.
Proof of claim 6.1.8. The first part of the proof is devoted to proving the following estimate: there exists C > 0 such that for any sequence (y ) of points in
M \ pi=1 Byi, (R0 i, ),
u (y ) u0 (y )
j1

n
2n
0
2 1
C
i, (y ) + j, Rj, (y )
+ o (1)
(6.1.122)
i=1
where
n
2n
2
dg (yi, , y)
0i, (y) = i,
(6.1.123)
and j, is as in (6.1.107). The definition of j, depends on the choice of some

1
. We fix such that claim 6.1.7 applies, and consider that Aj, and
0 < < n+2
j, , as defined in the proof of claim 6.1.7, are with respect to such an . Given
i = 1, . . . , j 1, we let

n
= y s.t. (y) ( 2 1)(12) Rj, (y)(2n)(1)
(6.1.124)
i,
i,
i,
j,
where i, is as in (6.1.93), i, is as in (6.1.94), and j, is as in (6.1.107). We
also let
i, .
j, = M \ j1
i=1
(6.1.125)
By point (P1) of Theorem 4.1, up to a subsequence, there exists > 0, 0 as

+, such that
sup
M \p
i=1 Byi, ( )
|u u0 | 0
(6.1.126)
132
CHAPTER 6
as +. Let k {1, . . . , j 1}, and let R () be such that (6.1.96)(6.1.98)

and (6.1.103) hold. We claim that, for large,
(6.1.127)
By (R () k, )
k,
k,
and
=
Byi, (R () i, )
k,
(6.1.128)
for all i = k. First we prove (6.1.127). Let (z ) be a sequence of points in

Byk, (R () k, ). We need to prove that, for large,
k, (z ) i, (z ) for all i {1, . . . , j 1} and
( n 1)(12)
(2n)(1)
Rj, (z )
.
k, (z ) j,2
We assume by contradiction that i, (z ) k, (z ) for some i = 1, . . . , j 1,
i = k. Since z Byk, (R () k, ), this gives that
n

( 2 1)(12) n2 1
(n2)(1)
= O i,
k,
dg (yi, , z )

(n2)(1)
.
= O max {i, ; k, }
Thus dg (yi, , z ) = O (max {i, ; k, }), and we get that
= O (max {i, ; k, }) + O (k, )
= O (max {i, ; k, }) .
Noting that this is in contradiction with (6.0.8), we have proved that z k, .
Assume now by contradiction that
( n 1)(12)
(2n)(1)
k, (z ) j,2
Rj, (z )
.
By (6.1.103), j, = O (j1, ). Since k j 1 and z Byk, (R () k, ),
the above equation gives that

n
( 2 1)(12) n2 1
(n2)(1)
Rj, (z )
= O j1,
k,

(n2)(1)
.
= O k,
Hence, Rj, (z ) = O (k, ). Let l {j, . . . , p} be such that, up to a subsequence,
Rj, (z ) = dg (yl, , z ). Then,
dg (yl, , yk, ) dg (yl, , z ) + dg (z , yk, )
= O (k, ) ,
, and (6.1.127) is proved.
and this is in contradiction with (6.0.8). Thus z
k,
Now we prove (6.1.128). We let i {1, . . . , p}, i = k, and assume by contradic . Then
tion that there exists a sequence (z ) of points in Byi, (R () i, )
k,
dg (z , yi, ) = O (i, ), and
( n 1)(12)
(2n)(1)
dg (yi, , z )
.
k, (z ) i,2
133
If i j 1, such an equation follows from the fact that z k, . If i j,

we write that j, j, i, and that Rj, (z ) dg (yi, , z ). With such an
equation we get that
n

( 2 1)(12) n2 1
(n2)(1)
= O k,
i,
dg (yk, , z )

(n2)(1)
,
= O max {i, ; k, }
and this leads to a contradiction as above. In particular (6.1.128) is proved.
We claim now that, for any k {j, . . . , p},
Byk, (R () j, )
j,
(6.1.129)
and that, for any k {1, . . . , j 1},

=
Byk, (R () k, )
j,
(6.1.130)
when is large. We prove (6.1.129). Let k {j, . . . , p}, and let (z ) be a

sequence of points in Byk, (R () j, ). We assume by contradiction that z is
. Hence we assume that there exists i {1, . . . , j 1} such
such that z
j,
that
( n 1)(12)
(2n)(1)
i, (z ) j,2
Rj, (z )
.
Since Rj, (z ) dg (yk, , z ) = O (j, ),
n 1 (12)

i, ( 2 )
(n2)(1)
(n2)(1)
dg (yi, , z )
Rj, (z )
j,
n

( 2 1)(12) n2 1
= O i,
j,
.
By (6.1.103), j, = O (j1, ), and since i j 1, j1, i, . As a
consequence, dg (yi, , z ) = O (i, ), and we can write that
= O (i, ) + O (j, )
= O (i, ) + O (j1, )
= O (i, ) .
Such an equation is is in contradiction with (6.0.8). This proves (6.1.129). An
easy remark is that (6.1.130) is a direct consequence of (6.1.127).
It follows from claim 6.1.7 and (6.1.128) that, for any k {1, . . . , j 1}, and
\By (R () k, ),
any y
k,
k,

2 1
2 1
(n+2)
C0 k, (y)
+ R (y)
(6.1.131)
u (y)
where C0 > 0 is independent of . Similarly, it follows from claim 6.1.7 and
\ p By (R () j, ),
(6.1.130) that, for any y
i,
j,
i=j
2 1
u (y)
( n2 +1)(12)
C0 j,
(n+2)(1)
Rj, (y)
(n+2)
+ R (y)
(6.1.132)
134
CHAPTER 6
where C0 > 0 is as above.

Now we prove that (6.1.122) holds. For that purpose, we let (y ) be a sequence
of points in M \ pi=1 Byi, (R0 i, ), and distinguish four cases.
Case 1. We assume that R (y ) as + for some > 0. Then, for
S as in Theorem 4.1, dg (y , S) as +, and point (P1) of Theorem 4.1
gives that (6.1.122) holds with C > 0.
Case 2. We assume that there exists k {1, . . . , p}, and R R0 , such that
dg (yk, , y )
R
k,
as +. By point (P1) of Theorem 4.1, we then get that
n
2
u (y ) = u (y0 yk )
lim k,
where yk = xNk is given by Theorem 4.1, u is given by (4.1.5), and, up to a

subsequence,
1
y0 = lim
exp1
yk, (y ) .
+ k,
Clearly, |y0 | R0 . Independently,
n

n
2n
2 1
2 1
lim k,
k,
= |y0 |2n .
dg (yk, , y )
+
It follows that
lim
u (y )
= |y0 |n2 u (y0 yk ) .
0k, (y )
(6.1.133)
Letting
D0 = 2 sup
sup |z|n2 u (z yk ) ,
(6.1.134)
k=1,...,p zRn
it is easily checked that (6.1.133) gives that

lim sup j1
+
u (y ) u0 (y )
0
i=1 i,
n
2 1
2n
(y ) + j, Rj, (y )
1
D0 .
2
In particular, (6.1.122) holds with C D0 .

Case 3. We assume that, for any k {1, . . . , p},
dg (yk, , y )
+
k,
as +, and that Rj, (y ) = O (j, ). First we prove that R (y ) =
Rj, (y ). Let l {j, . . . , p} be such that Rj, (y ) = dg (yl, , y ) = O (j, ).
We proceed by contradiction and assume that R (y ) = dg (yk, , y ) for some
k {1, . . . , j 1}. Then dg (yk, , y ) dg (yl, , y ), so that
dg (yk, , yl, ) dg (yk, , y ) + dg (y , yl, )
2dg (y , yl, )
= O (j, )
= O (j1, ) .
135
Since j 1 k, it follows from the above equation that dg (yk, , yl, ) = O (k, ),
and this is in contradiction with (6.0.8). Therefore,
R (y ) = Rj, (y ) = O (j, ) .
1
dg (yk, , y ) k,
Since
from (6.0.12) that
Hence,
+ as + for all k {1, . . . , p}, it follows

1 n
u (y ) u0 (y ) = o R (y ) 2 .

u (y ) u0 (y )

1 n
n2
= j, 2 R (y )
u (y ) u0 (y )

n 1
R (y ) 2
=o
j,
1 n
n2
j, 2 Rj, (y )
= o (1) .
In particular, (6.1.122) holds with C > 0.
Case 4. We assume that R (y ) 0 as +, that for any k {1, . . . , p},
dg (yk, , y )
+
k,
as +, and that
Rj, (y )
+
j,
as +. We let G be the Greens function of g +h . Thanks to the Greens
representation formula,

2 1
2 1
dvg
G (y , y) u (y)
u0 (y)
u (y ) u0 (y ) =
M
(6.1.135)
+
G (y , y) (h (y) h (y)) u0 (y) dvg .
M
From (6.0.1) and (6.1.80),
G (y , y) (h (y) h (y)) u0 (y) dvg = o (1) .
(6.1.136)
Independently, (6.1.80) and (6.1.126) give that

2 1
2 1
dvg = o (1) .
G (y , y) u (y)
u0 (y)
M \p
i=1 Byi, ( )
Inserting this equation and (6.1.136) into (6.1.135), we get that

u (y ) u0 (y )
p
i=1 Byi, ( )
2 1
G (y , y) u (y)
dvg + o (1) .
(6.1.137)
136
CHAPTER 6
Let k {1, . . . , j 1}. We set

1
p By ( ) s.t. dg (y , y) 1 dg (yk, , y ) ,
Bk,
= y
k,
i,
i=1
2

2
p By ( ) s.t. dg (y , y) < 1 dg (yk, , y ) ,
= y
Bk,
k,
i,
i=1
2
where is as in (6.1.126). Then
(p By
( ))
i=1
k,
i,
2 1
G (y , y) u (y)
2 1
1
Bk,
G (y , y) u (y)
dvg
2 1
2
Bk,
G (y , y) u (y)
dvg
(6.1.138)
dvg .
1
+ as +, for any y Byk, (R () k, ), and
Since dg (yk, , y ) k,
for large,
1
dg (y , y) dg (y , yk, ) .
2

1
Byk, (R () k, ) pi=1 Byi, ( ) Bk,
(6.1.139)
for large. From (6.1.80),

2n
G (y , y) C2 2n2 dg (yk, , y )
1
for all y Bk,
. It follows that, for large,
2 1
dvg
G (y , y) u (y)
p
Byk, (R()k, )(i=1 Byi, ( ))
2n
2 1
n2
C2 2
dg (yk, , y )
u (y)
dvg
Byk, (R()k, )
2n
C2 2n2 dg (yk, , y )
1

u 22 1 Volg Byk, (R () k, ) 2 .
Hence, for large,
2 1
Byk, (R()k, )(p

i=1 Byi, ( ))
D1 0k,
G (y , y) u (y)
dvg
(6.1.140)
(y )
where
D1 = 2n1 C2 2
n1
21
R () 2
(6.1.141)

n

( 2 +1)(12)
2 1
(n+2)(1)
(n+2)
u (y)
C0 k,
dg (yk, , y)
+ R (y)
137

1
\Byk, (R () k, ). Thus we can write with (6.1.80) that
for all y Bk,
2 1
G (y , y) u (y)
dvg
1 \B
Bk,
yk, (R()k, )
C0
(n+2)
1 \B
Bk,
yk, (R()k, )
G (y , y) R (y)
dvg
n
2n ( 2 +1)(12)
+C0 C2 2n2 dg (yk, , y )
k,
(n+2)(1)
dg (yk, , y)
dvg .
1 \B
Bk,
yk, (R()k, )
Noting that <
2
n+2 ,
it is easily checked that

(n+2)(1)
1 \B
Bk,
yk, (R()k, )
dg (yk, , y)
dvg
2n1
n(n+2)(1)
(R () k, )
(n + 2) (1 ) n
when is large. Independently, using Holders inequality and (6.1.80), and since
1
, we easily get that
< n+2
(n+2)
G (y , y) R (y)
dvg
1 \B
Bk,
yk, (R()k, )
G (y , y)
n
n1
n1
n
dvg
n1
n(n+2)
p
i=1 Byi, ( )
R (y)
dvg
= o (1) .
It follows that
2 1
1 \B
Bk,
yk, (R()k, )
D2 0k,
G (y , y) u (y)
dvg
(6.1.142)
(y ) + o (1)
where
D2 = C0 C2 2n1
n1
n(n+2)(1)
R ()
.
(n + 2) (1 ) n
Combining (6.1.140) and (6.1.142), and using (6.1.139), we get that
2 1
G (y , y) u (y)
dvg (D1 + D2 ) 0k, (y ) + o (1)
1
Bk,
2
, we have that
for large. Independently, for y in Bk,
dg (yk, , y) dg (yk, , y ) dg (y , y)
1
dg (yk, , y ) dg (yk, , y )
2
1
dg (yk, , y ) .
2
(6.1.143)
(6.1.144)
138
CHAPTER 6
2
It follows, from (6.1.131) and (6.1.139), that for any y Bk,
,
2 1
u (y)
C0
( n2 +1)(12)
k,
dg (yk, , y )
2
(n+2)(1)
(n+2)
+ R (y)
We can then write that
2 1
G (y , y) u (y)
dvg
2
Bk,
C0
(n+2)
2
Bk,
G (y , y) R (y)
( n2 +1)(12)
+C0 k,
dg (yk, , y )
2
dvg
(n+2)(1)
2
Bk,
G (y , y) dvg .
As above, it is easily checked that
(n+2)
G (y , y) R (y)
dvg = o (1) .
2
Bk,
Independently, we have using (6.1.80) that
2n
G (y , y) dvg C2
dg (y , y)
dvg
2
Bk,
2
Bk,
C2
By ( 12 dg (yk, ,y ))
dg (y , y)
2n
dvg

2
= O dg (yk, , y ) .
Hence,
2 1
2
Bk,
G (y , y) u (y)
=O

0k,

(y )
dvg
k,
dg (yk, , y )
2(n+2)
+ o (1) .
1
2
Noting that < n+2
, and that dg (yk, , y ) k,
+ as +, it follows
that

2 1
G (y , y) u (y)
dvg = o 0k, (y ) + o (1) .
(6.1.145)
2
Bk,
Coming back to (6.1.138), we obtain with (6.1.144) and (6.1.145) that, for any
k {1, . . . , j 1},
2 1
dvg
G (y , y) u (y)
p
(
(6.1.146)
i=1 Byi, ( ))
k,
(D1 + D2 + o (1)) 0k, (y ) + o (1)
139
where D1 is as in (6.1.141) and D2 is as in (6.1.143).

Given k {j, . . . , p}, we now let

\ p By (R () j, ) s.t. Rj, (y) = dg (yk, , y)
k, = y
j,
i,
i=j
and

1
1
= y k, pi=1 Byi, ( ) s.t. dg (y , y) Rj, (y ) ,
Bk,
2

1
2
= y k, pi=1 Byi, ( ) s.t. dg (y , y) < Rj, (y ) .
Bk,
2
We write that
2 1
(p By
( ))
i=1
j,
i,
G (y , y) u (y)

k=j
dvg
2 1
k, (
p
i=1 Byi, ( )
G (y , y) u (y)
2 1
p
i=j Byi, (R()j, )
G (y , y) u (y)
dvg
(6.1.147)
dvg .
1
Since Rj, (y ) j,
+ as +, we get that
dg (y , y)
1
Rj, (y )
2
for all y pi=j Byi, (R () j, ) when is large. Then, by (6.1.80), we get that
for large
2 1
G (y , y) u (y)
dvg
p
i=j Byi, (R()j, )
2n
C2 2n2 Rj, (y )
2 1
p
i=j Byi, (R()j, )
2n
C2 2n2 Rj, (y )
u (y)
dvg
1

u 22 1 Volg pi=j Byi, (R () j, ) 2
2n
2
(p j + 1) D1 j,
Rj, (y )
where D1 is as in (6.1.141). Coming back to (6.1.147), we get that
2 1
dvg
G (y , y) u (y)
p
j, (i=1 Byi, ( ))
n
2n
2
(p j + 1) D1 j,
Rj, (y )
p

2 1
+
dvg .
G (y , y) u (y)
p
k=j k, (i=1 Byi, ( ))
(6.1.148)
140
CHAPTER 6
Let k {j, . . . , p}. We write that
k, (p
i=1 Byi, ( ))
2 1
G (y , y) u (y)
2 1
1
Bk,
G (y , y) u (y)
dvg
2 1
2
Bk,
dvg
G (y , y) u (y)
(6.1.149)
dvg .
1
, we have from (6.1.80) that
In Bk,
2n
G (y , y) C2 dg (y , y)
2n
C2 2n2 Rj, (y )
and using (6.1.132) and the definition of k, we can write that

n

( +1)(12)
2 1
(n+2)(1)
(n+2)
u (y)
.
C0 j,2
dg (yk, , y)
+ R (y)
Thus
2 1
1
Bk,
G (y , y) u (y)
C0 C2 2
n2
2n
Rj, (y )
+C0
dvg
( n2 +1)(12)
j,
(n+2)(1)
1
Bk,
(n+2)
1
Bk,
G (y , y) R (y)
dg (yk, , y)
dvg
dvg .
As above, one easily gets that
(n+2)
G (y , y) R (y)
dvg = o (1) .
1
Bk,
Since <
2
n+2 ,
we also have that, for large,
(n+2)(1)
dg (yk, , y)
dvg
1
Bk,
(n+2)(1)
M \Byk, (R()j, )
dg (yk, , y)
dvg
2n1
n(n+2)(1)
(R () j, )
.
(n + 2) (1 ) n
Thus we have that
n
2 1
2n
2 1
G (y , y) u (y)
dvg D2 j,
Rj, (y )
+ o (1)
1
Bk,
2
, then
where D2 is as in (6.1.143). Independently, if y is in Bk,
Rj, (y) = dg (yk, , y)
1
Rj, (y ) .
2
(6.1.150)
141
2
,
From (6.1.132), it follows that, for y Bk,

(n+2)(1)

n
( 2 +1)(12) Rj, (y )
2 1
(n+2)
C0 j,
+ R (y)
.
u (y)
2
Then,
2 1
2
Bk,
G (y , y) u (y)
dvg
n

( 2 +1)(12)
(n+2)(1)
= O j,
Rj, (y )
2
Bk,
+O
(n+2)
2
Bk,
G (y , y) R (y)
( n2 +1)(12)
= O j,
so that
Rj, (y )
2
Bk,
G (y , y) dvg + o(1) .
G (y , y) dvg C2
By
2 1
2
Bk,
G (y , y) u (y)

n
2 1
2n
= O j, Rj, (y )
Since <
dvg
(n+2)(1)
From (6.1.80),
2
Bk,
G (y , y) dvg
j, (y )
2
dg (y , y)
2n
dvg
dvg
Rj, (y )
j,
(n+2)2
+ o (1) .
1
and Rj, (y ) j,
+ as +, we get that

n

2 1
2n
2 1
+ o (1) .
G (y , y) u (y)
dvg = o j,
Rj, (y )
2
n+2
2
Bk,
Coming back to (6.1.149), from (6.1.150), this implies that
2 1
dvg
G (y , y) u (y)
k, (p
i=1 Byi, ( ))
n
2n
2
(D2 + o (1)) j,
Rj, (y )
+ o (1) .
Coming back to (6.1.148) we then get that
2 1
dvg
G (y , y) u (y)
(p By
( ))
i=1
j,
i,
(6.1.151)

n
2n
2 1
(p j + 1) (D1 + D2 ) + o (1) j, Rj, (y )
+ o (1) .
, we get with (6.1.137), (6.1.146), and (6.1.151) that

Noting that M = jk=1
k,
u (y ) u0 (y )
(D1 + D2 + o (1))
j1
0k, (y )
k=1
n
2n
2
+ (p j + 1) (D1 + D2 + o (1)) j,
Rj, (y )
+ o (1) .
142
CHAPTER 6
In particular, (6.1.122) holds with C > p (D1 + D2 ).

From cases 14, for any sequence (y ) in M \ pj=1 Byj, (R0 j, ), equation
(6.1.122) holds with C D0 + p (D1 + D2 ), where D0 is as in (6.1.134), D1 is
as in (6.1.141), and D2 is as in (6.1.143). In order to prove that (6.1.90)j holds,
and thus claim 6.1.8, it remains to prove that
j, = O (j, ) .
(6.1.152)
By (6.1.107), this reduces to proving that Aj, = O (1) where Aj, is given by
(6.1.101). We proceed by contradiction and assume that
Aj, +
(6.1.153)
as +. By the definition (6.1.101) of Aj, , up to a subsequence, there exist
k {1, . . . , j 1} and y k, such that
n
1 (12)
(2n)(1)
(Aj, j, )( 2 )
Rj, (y )
= k, (y ) .
Clearly, (6.1.153) implies that j, = Aj, j, . Hence,
( n 1)(12)
(2n)(1)
Rj, (y )
= k, (y ) .
(6.1.154)
j,2
Since y k, , k, (y ) i, (y ) for all i {1, . . . , j 1}, and

1

n
D () 2 2
1
dg (yk, , y ) 2
u (y ) u0 (y ) .
100
Using (6.1.122), and the definition (6.1.102) of k, , we can then write with
(6.1.154) that
1

D () 2 2
100

j1

n
n
1
1
2n
2
Cdg (yk, , y ) 2
0i, (y ) + j,
Rj, (y )
+ o (1)
i=1

n
Cdg (yk, , y ) 2

+
Rj, (y )
j,
Rj, (y )
j,
(n2)
R ()
j1
i, (y )
i=1
(n2)
( n2 1)(12)
j,

(2n)(1)
Rj, (y )
+ o (1)
Cdg (yk, , y )

n
2 1
k,
(n2)
(n2)
(y ) (j 1) R ()

+ o (1) .
Therefore,
1

1 n (12)
dg (yk, , y ) ( 2 )
D () 2 2
C
100
k,

(n2)
Rj, (y )
(n2)
(j 1) R ()
+
+ o (1) .
j,
(6.1.155)
143
Let us now assume that dg (yk, , y ) = O (k, ). Coming back to (6.1.154), we

get that
n 1 (12)

j, ( 2 )
(n2)(1)
(n2)(1)
=
dg (yk, , y )
Rj, (y )
k,
n

( 2 1)(12) n2 1
= O j,
k,
.
By (6.1.103), j, = O (j1, ). By (6.0.7), since k j 1, we also have that
j1, k, . It follows that
Rj, (y ) = O (k, ) .
Let l {j, . . . , p} be such that Rj, (y ) = dg (yl, , y ). Then,
dg (yk, , yl, ) dg (yk, , y ) + dg (y , yl, )
= O (k, )
and this is in contradiction with (6.0.8). Therefore,
dg (yk, , y )
+
k,
as +. As a consequence of (6.1.155) we then get that Rj, (y ) is such
that Rj, (y ) = O (j, ). Coming back to (6.1.154), it follows that

n 1 (12)
k, ( 2 )
(n2)(1)
(n2)(1)
dg (yk, , y )
=
Rj, (y )
j,

n
( 2 1)(12) n2 1
= O k,
j,

n
( 2 1)(12) n2 1
j1,
= O k,

(n2)(1)
= O k,
1
+ as +. In parand this is a contradiction since dg (yk, , y ) k,
ticular, (6.1.153) is absurd and (6.1.152) is proved. Together with (6.1.122), this
ends the proof of claim 6.1.8.
2

we then get that (6.1.90)j is true for all j = 1, . . . , p. In particular, (6.1.90)p is
true. Noting that equations (6.1.67) and (6.1.90)p are the same, (6.1.67) is proved.
6.2 A FUNDAMENTAL ESTIMATE: 2
We prove the fundamental estimate with respect to all blow-up points, namely, the
upper estimate of Theorem 6.1. In other words, we prove that there exist C > 0
and a sequence ( ) of positive real numbers converging to 0 as + such
that
N

u (x) (1 + ) u0 (x) + C
i, (x)
(6.2.1)
i=1
144
CHAPTER 6
for all x M and all , where
i, (x) =
i,
2i, +
dg (xi, ,x)2
n(n2)
n2
2
It follows from section 6.1 [we refer to (6.1.1) and (6.1.67)] that (6.2.1) holds
for all x M \ pj=1 Byj, (R0 j, ). It remains to prove that (6.2.1) holds also in
Byj, (R0 j, ) for all j {1, . . . , p}. We fix j0 {1, . . . , p}, and let
q, = xNj0 +1 1, ,
x
1, = xNj0 +1, , . . . , x
q, = Nj0 +1 1, .
1, = Nj0 +1, , . . . ,
It easily follows from point (P1) of Theorem 4.1 that (6.2.1) holds in the ball
Byj0 , (R0 j0 , ) if there exists C > 0 such that
q

n 1
1 n
2n
2
2
i, dg (
xi, , x)
+ j0 ,
u (x) C
(6.2.2)
i=1
x1, , . . . , x
q, } and all . Without loss of generalfor all x Byj0 , (R0 j0 , ) \ {
ity, we assume in what follows that the set {
xi, }i=1,...,q is not empty. Otherwise,
(6.2.2) is a direct consequence of Theorem 4.1.
We rearrange the x
i, s by a process analogous to that used in the introduction
of this chapter. We let x
1, be such that
1,
i, for all i {1, . . . , q}. We
q2 1 the x
i, s, i = 2, . . . , q, which are such that, up to
denote then by x
2, , . . . , x
a subsequence,
lim
dg (
x1, , x
i, )
< + .
1,
q2 , be such that
q2 ,
i, for all i {q2 , . . . , q}. We
We let q1 = 1, and let x
denote then by x
q2 +1, , . . . , x
q3 1, the x
i, s, q2 + 1 i q, which are such
that, up to a subsequence,
lim
dg (
xq2 , , x
i, )
< + .
q2 ,
Going on with this process, we obtain a sequence of points

q1 , , x
2, , . . . , x
q2 , , x
q2 +1, , . . . , x
q3 , , . . . , x
q, = x
qr+1 1, .
x
1, = x
For i {1, . . . , r}, we let
qi ,
yi, = x
and
i, =
qi , .
The above construction gives that

1, r,
(6.2.3)
and that for any i, j {1, . . . , r}, i = j,

yi, , yj, )
dg (
+
i,
(6.2.4)
145
0 > 0 be such that for any j {1, . . . , r} and for any

as +. We let R
i {qj + 1, . . . , qj+1 1},
0
dg (
yj, , x
i, )
R
.
j,
4
lim sup
+
We have that

SNj0 =
lim
j0 ,
exp1
yj0 ,
(6.2.5)

(
yj, ) j = 1, . . . , r
where SNj0 is as in (4.1.2). We claim now that there exists C > 0 such that
r

n
n
1
1
2n
2
j,
dg (
yj, , y)
+ j0 ,2
u (y) C
(6.2.6)
j=1
0 j, ) and all . We prove this claim

for all y Byj0 , (R0 j0 , )\ rj=1 Byj, (R
in what follows. We divide the proof into several steps, following what we did
when proving (6.1.67). Noting that the arguments developed below do not use any
specific property of yj0 , , but only the properties satisfied by any concentration
point and the above construction, we easily get by finite induction that an equation
like (6.2.6) implies (6.2.2). In particular, the proof of (6.2.1) reduces to the proof
of (6.2.6). The first step in the proof of (6.2.6) is the following:
j, ),
C LAIM 6.2.1. Given R > 0, for any y Byj0 , (R0 j0 , ) \rj=1 Byj, (R
and for any ,
(y) 2 1 |u (y) vN , (y) | R,
R
j0
n
where
(y) = min dg (
yj, , y) ,
R
j=1,...,r
where lim
lim R, = 0, and vNj0 , is as in Theorem 4.1.
R+ +
Proof of claim 6.2.1. We proceed by contradiction and assume that claim 6.2.1
does not hold. Then there exists a sequence (xN +1, ) in Byj0 , (R0 j0 , ) such
that, for any j {1, . . . , r},
yj, , xN +1, )
dg (
+
j,
(6.2.7)
as +, and such that

(xN +1, ) 2 1 |u (xN +1, ) vN , (xN +1, ) | (40 ) 2 1
R
j0
n
(6.2.8)
for some 0 > 0. We let

1 n
2
u (xN +1, ) = N +1,
(6.2.9)
and prove that Theorem 4.1 continues to hold if we add xN +1, and N +1, to the
xi, s and i, s, i = 1, . . . N . Noting that this contradicts the maximality of N , as
146
CHAPTER 6
assumed in the introduction of this chapter, claim 6.2.1 will be proved. We follow
the lines of step 3 of the proof of Theorem 4.1. We claim that
lim 1
+ j0 ,
exp1
yj0 , (xN +1, ) SNj0
(6.2.10)
where SNj0 is defined by (4.1.2). If (6.2.10) is false, point (P1) of Theorem 4.1
gives that

1 n
|u (xN +1, ) vNj0 , (xN +1, ) | = o j0 ,2 .
(xN +1, ) = O (j , ), we get that
Noting that this contradicts (6.2.8) since R
0
(6.2.10) holds. From (6.2.10),
(xN +1, ) = o (j , ) .
R
(6.2.11)
0
1 n2
Since vNj0 , (xN +1, ) = O j0 , , we get by combining (6.2.8), (6.2.9) and
(6.2.11) that
(xN +1, )
R
20
N +1,
(6.2.12)
N +1, = o (j0 , ) .
(6.2.13)
for large, and that

In particular, N +1, 0 as +. Now we consider the following assertions:
(D1) For any i = 1, . . . , N ,
dg (xi, , xN +1, )
+
min {i, , N +1, }
as +.
(D2) There exists 0 > 0 and xN +1 Rn such that
lim N +1, (x)uN +1, (x) = u (x xN +1 )
2
(Rn \SN +1 ), and strongly in C 2 (B0 (0 )),
weakly in D12 (Rn ), strongly in Cloc
where N +1, is as in Chapter 4, SN +1 is as in (4.1.2), uN +1, is as in (4.1.3),
and u is as in (4.1.5).
(D3) The energy E, given by E(u) = u2 , satisfies

2
N
+1

vi,
lim E u u0
+
i=1
2
lim E (u )
+
E(u0 )2 (N + 1) 2min
where vi, = vi, (ui ) is as in (4.1.4), ui (x) = u (x xi ), u is as in (4.1.5), and

xi is as in (P2) and (D2).
As is easily checked, the proof of claim 6.2.1 reduces to the proof of assertions
(D1)(D3). Indeed, (D1) implies that (4.1.6) is satisfied by the xi, s and i, s,
i = 1, . . . , N + 1. Then, point (P1) of Theorem 4.1 with respect to the xi, s and
147
i, s, i = 1, . . . , N + 1, is a consequence of (D2). One may note here that the set

S is unchanged when we add the sequence (xN +1, ). Independently, point (P2)
of Theorem 4.1 obviously continues to hold when we add (sequences of) blow-up
points. Finally, (D3) implies that point (P3) of Theorem 4.1 holds for the xi, s and
i, s, i = 1, . . . , N + 1. Hence, if (D1)(D3) hold, then Theorem 4.1 continues
to hold if we add xN +1, and N +1, to the xi, s and i, s, i = 1, . . . N . This
contradicts the maximality of N and proves claim 6.2.1.
We prove (D1). First we consider the case where i {Nj0 , . . . , Nj0 +1 1}.
Then, by construction of the yj, s,
dg (xi, , yj0 , )
+
j0 ,
as +. Since dg (xN +1, , yj0 , ) = O (j0 , ) and using (6.2.13), this leads
to
dg (xi, , xN +1, )
+
N +1,
as +. Now we consider the
case
i = Nj0 . It is easily checked (see the proof
of Theorem 4.1) that SNj0 B0 20 is empty for some 0 > 0. From (6.2.10), we
can then write that dg (yj0 , , xN +1, ) 40 j0 , for large, and it follows from
(6.2.13) that
dg (yj0 , , xN +1, )
+
N +1,
as +. Finally, we consider the case where i {Nj0 + 1, . . . , Nj0 +1 1}.
k, for some k {1, . . . , q}. We let j {1, . . . , r} be such that
Then xi, = x
k {qj , . . . , qj+1 1}. By (6.2.5), we can write that
dg (
xk, , xN +1, ) dg (xN +1, , yj, ) dg (
xk, , yj, )
0
R
dg (xN +1, , yj, )
j,
2
d (
x
,x
for large. Then (6.2.7) implies that g k,j,N +1, + as +. In

particular, since
k, j, by construction,
dg (
xk, , xN +1, )
+
k,
as +. Summarizing, from the above three cases, (D1) is proved. Now we
prove (D2). Given x B0 (1
N +1, ), the Euclidean ball of center 0 and radius
,
we
let
1
N +1,

n
1
uN +1, (x) = N2 +1, u expxN +1, (N +1, x) ,
gN +1, (x) = expxN +1, g (N +1, x) ,

hN +1, (x) = h expxN +1, (N +1, x) .
(6.2.14)
148
CHAPTER 6
Since N +1, 0 as +,
2
lim gN +1, = in Cloc
(Rn ) .
(6.2.15)
We also have that gN +1, is controlled on both sides by in the sense of bilinear
forms. It is easily checked that
uN +1, (0) = 1
(6.2.16)
and that
2 1
gN +1, uN +1, + 2N +1, hN +1, uN +1, = uN
+1,
(6.2.17)
in B0 (1
N +1, ). We let

exp1
xN +1, (xi, )
SN +1 =
lim
, xi, BxN +1, () , 1 i N
+
N +1,
where, up to a subsequence, the limits are assumed to exist. By (6.2.12) [see also
the proof of (D1)]

3
SN +1 B0
0 = .
(6.2.18)
2
Let R > 0 and let (x ) be a sequence of points in B0 (R) such that
1
R
where d is the Euclidean distance. We can write with (6.2.15) that

N +1,
min dg xi, , expxN +1, (N +1, x )
i=1,...,N
2R
d (x , SN +1 )
(6.2.19)
for large. Let y = expxN +1, (N +1, x ). It follows from point (P2) of Theorem 4.1 that
uN +1, (x ) (2R) 2
(2R) 2
min dg (xi, , y ) 2
i=1,...,N
u (y )
for some C > 0 independent of . Thus (uN +1, ) is locally uniformly bounded
in Rn \SN +1 . By standard elliptic theory, and (6.0.1), (6.2.14), (6.2.15), and
(6.2.17), we then get that, up to a subsequence,
lim N +1, uN +1, = uN +1
(6.2.20)
2
(Rn \SN +1 ). We also have that N +1, uN +1, uN +1 weakly in the
in Cloc
space D12 (Rn ) as +. Moreover, uN +1 verifies that
2 1
n
uN +1 = uN
+1 in R
and uN +1 (0) = 1 thanks to (6.2.16). By Caffarelli-Gidas-Spruck [17],

n2

2
N +1
uN +1 (x) =
1
1 + n(n2)
2N +1 |x xN +1 |2
149
for some N +1 > 0 and some xN +1 Rn such that

N +1 = 1 +
1
2 |xN +1 |2 .
n(n 2) N +1
Up to changing N +1, into 1

N +1 N +1, and xN +1 into N +1 xN +1 , (D2) is
proved. Now, it remains to prove (D3). Let R > 0. For any > 0, there exists C
independent of R and such that

2
N
+1

vi, dvg
u u0

M \BxN +1, (RN +1, )
i=1

2
N

(1 + )
vi, dvg
u u0

M \BxN +1, (RN +1, )
i=1
2
+C
vN
+1, dvg .
M \BxN +1, (RN +1, )
Noting that
lim
lim
R+ +
2
vN
+1, dvg = 0 ,
M \BxN +1, (RN +1, )
we get that

2
N
+1

0
vi, dvg
u u

M \BxN +1, (RN +1, )
i=1

2
N

vi, dvg + R,
u u0

M \Bx
(RN +1, )
i=1
N +1,
where lim
(6.2.21)
lim R, = 0. Hence,
R+ +
2

N
+1

0
vi, dvg
u u

M
i=1
2

N

vi, dvg + IR () + R,
u u0

M
(6.2.22)
i=1
where
2

N
+1

vi, dvg
IR () =
u u0

BxN +1, (RN +1, )
i=1
2

N

vi, dvg .
u u0

Bx
(RN +1, )
N +1,
i=1
(6.2.23)
150
CHAPTER 6
We have that

2
N
+1

vi, dvgN +1,
IR () =
uN +1, u0

B0 (R)
i=1

2
N

vi, dvgN +1,

uN +1, u0

B0 (R)
i=1
where

n
1
vi, (x) = N2 +1, vi, expxN +1, (N +1, x)
and

n
1
u0 (x) = N2 +1, u0 expxN +1, (N +1, x) .
We let VR be given by
VR =

x Rn s.t. d (x, SN +1 )
Noting that
lim
lim
R+ +
VR
1
R

.
2
vN
+1, dvgN +1, = 0 ,
similar arguments to those used to prove (6.2.21) give that

2
N
+1

0
vi, dvgN +1,
IR ()
uN +1, u

B0 (R)\VR
i=1
(6.2.24)

2
N

vi, dvgN +1, + R, .

uN +1, u0

B0 (R)\VR
i=1
Given i = 1, . . . , N , either dg (xi, , xN +1, ) = O (N +1, ) or

dg (xi, , xN +1, )
+
N +1,
as +. Since (4.1.6) is satisfied by the xi, s and i, s, i = 1, . . . , N + 1,
+1,
+ as +. Hence, given i =
we get in the first case that Ni,
1, . . . , N , either
dg (xi, , xN +1, )
+
(6.2.25)
N +1,
as +, or
N +1,
+
(6.2.26)
i,
as +. If i {1, . . . , N } is such that (6.2.25) holds, then

2n
2
vi,
dvg = O ni, dg (xi, , xN +1, )
BxN +1, (RN +1, )

Volg BxN +1, (RN +1, )

2n
= O ni, nN +1, dg (xi, , xN +1, )
n

i,
=o
.
dg (xi, , xN +1, )
151
yj, ),
From (6.2.7) and the construction of (yj, ) and (
i, = O (dg (xi, , xN +1, ))
for all i = 1, . . . , N . Thus,
if i {1, . . . , N } is such that (6.2.25) holds, then

2
vi,
dvg = o (1)
BxN +1, (RN +1, )
and we get that
2
vi,
dvgN +1, 0
B0 (R)\VR
as +. Let now i {1, . . . , N } be such that (6.2.26) holds and (6.2.25)

does not hold. Let also x B0 (R) \VR . Then, for sufficiently large,

1
dg xi, , expxN +1, (N +1, x)
N +1,
2R
and it follows from the expression of vi, that

n2 1
i,

vi, L (B0 (R)\VR ) CR
N +1,
where CR > 0 is independent
of
.
In
particular,
2
vi,
dvgN +1, 0
B0 (R)\VR
as +. Summarizing,
we proved that for any i = 1, . . . , N ,

2
vi,
dvgN +1, 0
B0 (R)\VR
(6.2.27)
as +. Since point (P1) of Theorem 4.1 holds for xN +1, and N +1, ,

|uN +1, vN +1, |2 dvgN +1, 0
(6.2.28)
as +, and
lim
B0 (R)\VR
lim
R+ +
u2N +1, dvgN +1, = 2min .

B0 (R)\VR
From (6.2.27) and (6.2.28),

2
N

0
vi, dvgN +1, 0
uN +1, vN +1, u

B0 (R)\VR
i=1
as +. Independently, from (6.2.27) and (6.2.29),

2
N

vi, dvgN +1,
uN +1, u0

B0 (R)\VR
i=1
u2N +1, dvgN +1,

B0 (R)\VR
2 1
0
2
uN
+1, u dvgN +1,
B0 (R)\VR
2

i=1

2min
B0 (R)\VR
+ R,
2 1
uN
i, dvgN +1,
+1, v
(6.2.29)
152
CHAPTER 6
where lim
lim R, = 0. Coming back to (6.2.24), we then get that
R+ +
IR () 2min + R,
and using (6.2.22), since R is arbitrary, we get that (D3) is proved. Therefore,
(D1)(D3) hold. As already mentioned, this is in contradiction with the maximality
of N in Theorem 4.1. In particular, claim 6.2.1 is proved.
2
Going on with the proof of (6.2.6), mimicking what was done in section 6.1, we
now prove the following:
0 and C () > 0 such
C LAIM 6.2.2. For any 0 < < 12 , there exist R () R
that
n

( 2 1)(12)
(1 n )(12)
(2n)(1)
(2n)
u (y) C () 1,
+ j0 , 2
R (y)
R (y)
is as in
for all y Byj0 , (R0 j0 , ) \ rj=1 Byj, (R () j, ) and all , where R
claim 6.2.1.
Proof of claim 6.2.2. Let 0 < < 12 . Also let x0 SNj0 , where SNj0 is as in
Theorem 4.1, and x = expyj0 , (j0 , x0 ). We set
D () =
(1 ) C12
4r
C22
(6.2.30)
and (6.1.3). First we claim that there exist

where C1 and C2 are given by (6.1.2)

0 and 0 < () 1 d x0 , SN \ {x0 } such that
R () R
j0
2
(y)2 u (y)2
R
min {D () ; D (1 )}
(6.2.31)
for all y Bx ( () j0 , ) \
Byj, (R () j, ) and large. In order to
prove (6.2.31), we let (y ) be a sequence of points such that
rj=1
j, )
y Bx (j0 , ) \ rj=1 Byj, (R
for all . Using claim 6.2.1, we can write that
n
n
(y ) 2 1 u (y ) R, + R
(y ) 2 1 vN
R
j0 ,
where lim
(y )
lim R, = 0. Thus,
R+ +
(y ) 2 1 u (y ) R, + (2j , ) 2 1 vN , (y )
R
0
j0
n
.
R, + (2) 2
0 sufficiently large, this proves
Choosing > 0 sufficiently small and R R
(6.2.31). Now we consider the linear operator L given by
2 2
L (u) = g u + h u u
u.
This operator satisfies the maximum principle by [9]. We let h0 C 0, (M ) be as

in the introduction of section 6.1, and G be the Greens function of g + h0 . We
set
r
r

1
Gj, (y)
and H2 (y) =
Gj, (y)
H1 (y) =
j=1
j=1
153
yj, , y). Direct computations lead to

where Gj, (y) = G (

r

1
2 2
1
Gj, (y)
h (y) (1 ) h0 (y) u (y)
L H (y) =
j=1
+ (1 )
|Gj, (y)|2
2
Gj, (y)
and to
L H2 (y) =
r

2 2
Gj, (y)
h (y) h0 (y) u (y)
j=1
+ (1 )
|Gj, (y)|2
Gj, (y)
in M \ {
y1, , . . . , yr, }. Let
y Bx ( () j0 , ) \ rj=1 Byj, (R () j, ) .
By (6.1.2),
|Gj, (y )|2
2
Gj, (y )
C12
dg (
yj, , y )
for large. By (6.2.31), we also have that

2 2
u (y )
(y )2
min {D () ; D (1 )} R
for large. Let 1 l r be such that, up to a subsequence, we have that

(y ). Thanks to (6.1.3) and the two equations above, we can
yl, , y ) = R
dg (
write that, for large,
L H1 (y )

r

1
(y )2
Gj, (y )
h (y ) (1 ) h0 (y ) D (1 ) R
j=1
+ (1 )
C12

2
dg (
yj, , y )
C12
(2n)(1)2
dg (
yl, , y )
C21
r
r

1
1
(y )2
D (1 ) R
Gj, (y )
O
Gj, (y )
(1 )
j=1
j=1
C12
(2n)(1)2
(1 ) 1
R (y )
C2
(y )(2n)(1)2
rD (1 ) C21 R

(y )(2n)(1)
O R
(y )
2rD (1 ) C21 R
(2n)(1)2
154
CHAPTER 6
Similarly, we also have that

(y)
L H2 (y ) 2rD () C2 R
(2n)2
for large. Hence, when is large,

L H1 0 and L H2 0
in Bx ( () j0 , ) \
since
rj=1
(6.2.32)
Byj, (R () j, ). From point (P1) of Theorem 4.1,
()

1
d x0 , SNj0 \ {x0 } ,
2
we can write that

1 n
u 2j0 ,2 on Bx ( () j0 , )
0 , we also have that
for large. Since R () R
1 n
u 2
j, 2 on Byj, (R () j, )
for all j {1, . . . , r} and large. By (6.1.3) and (6.2.3) we then get the existence
of C () > 0 such that
n

( 2 1)(12) 1
(1 n2 )(12) 2
u C () 1,
H + j0 ,
H

on Bx ( () j0 , ) \ rj=1 Byj, (R () j, ) . From (6.2.32), the above equation, and the fact that L u = 0 in M , we can apply the maximum principle. It
follows that
n

( 2 1)(12) 1
(1 n )(12) 2
H + j0 , 2
H
u C () 1,
in Bx ( () j0 , ) \ rj=1 Byj, (R () j, ). In particular, using (6.1.3), this
proves that the estimate of claim 6.2.2 holds in

Bx ( () j0 , ) \ rj=1 Byj, (R () j, ) .
x0 SNj
Noting that, with a possibly larger C (), point (P1) of Theorem 4.1 gives that this
estimate also holds in
Byj0 , (R0 j0 , ) \ x0 SNj Bx ( () j0 , ) ,
0
this ends the proof of claim 6.2.2.

Our next claim is the following:
C LAIM 6.2.3. There exist C > 0 and a sequence ( ) of positive real numbers
converging to 0 as + such that
n
1 n
2
(y)
+ j0 ,2
u (y) vNj0 , (y) C 1,
R

0 j, and all , where vN , is as
for all y Byj0 , (R0 j0 , ) \ rj=1 Byj, R
j0
in Theorem 4.1, and R is as in claim 6.2.1.
2n
155
Proof of claim 6.2.3. We need to prove that there exists C

> 0 such
that for any
0 j, ,
sequence (y ) of points in Byj0 , (R0 j0 , ) \ rj=1 Byj, R

n
n
1
2n
2 1
0 . (6.2.33)
R (y )
lim sup j20 , u (y ) vNj0 , (y ) C 1,
+
We study the quantity involved in (6.2.33) and distinguish four cases.

Case 1. We assume that there exists > 0 such that, up to a subsequence,
as +. Then it follows from point (P1) of Theorem 4.1 that

1 n
u (y ) vNj0 , (y ) = o j0 ,2 .
(y )
R
j0 ,
Hence, (6.2.33) holds with C 0.

Case 2. We assume that there exists k {1, . . . , r} such that, up to a subsed (
y
,y )
0 . By (6.2.5) and point
quence, g k,
R as + for some R R
k,
(P1) of Theorem 4.1, we then get that, up to a subsequence,
n
1
lim 2 u
+ k,
(y ) = u (y0 yk )
where, up to a subsequence,
1
exp1
yk, (y )
k,
is given by Theorem 4.1. Noting that
dg (
yk, , y )
lim
= |y0 | ,
+
k,
y0 = lim
and yk = xNj0 +qk
we obtain that
lim
u (y )
n
2 1
2n
k, dg (
yk, , y )
= |y0 |n2 u (y0 yk ) .
From (6.2.3), it follows that

lim sup
+
u (y )
D0
(y )2n
R
n
2 1
1,
where
D0 =
sup
sup |z|n2 u (z yj )
j=1,...,r zRn
(6.2.34)
and yj = xNj0 +qj is as in Theorem 4.1. This gives that (6.2.33) holds with C >
D0 .
d (
y
,y )
Case 3. We assume that g j,

+ as + for all j {1, . . . , r}
j,
(y ) = O (
1, ). From claim 6.2.1,
and that R
1
n2
1, 2 R
|u (y ) vNj0 , (y ) |
(y )
n2 1

(y )
n
R
(y ) 2 1 |u (y ) vN , (y ) |
R
=
j0
1,

n
(y ) 2 1 |u (y ) vN , (y ) |
=O R
j0
n
= o (1) .
156
CHAPTER 6
Hence, (6.2.33) holds with C > 0.

(y ) = o (j , ) and that R
(y ) 1 + as
Case 4. We assume that R
0
1,
+. We let ( ) be a sequence of smooth functions with compact support in
Byj0 , (2R0 j0 , ) such that 1 in Byj0 , (R0 j0 , ) and

= O j1
, g = O j2
.
(6.2.35)
0 ,
0 ,
Let G be the Greens function of g + h . We write with the Greens representation formula that
u (y ) vNj0 , (y )

=
G (y , y) g u vNj0 , (y) dvg
Byj , (2R0 j0 , )
0

+
G (y , y) h (y) (y) u (y) vNj0 , (y) dvg .
Byj , (2R0 j0 , )
0
(6.2.36)
From (6.0.9), (6.2.35), and point (P1) of Theorem 4.1,

1 n
g u vNj0 , (y) = o j0 , 2
in Byj0 , (2R0 j0 , ) \Byj0 , (R0 j0 , ). Thus we can write that

Byj , (2R0 j0 , )\Byj , (R0 j0 , )
0
0

n
1 2
= o j0 ,
G (y , y) dvg .
Byj , (2R0 j0 , )
0
By (6.1.80), we have that

2n
dvg
G (y , y) dvg = O
dg (y , y)
Byj , (2R0 j0 , )
Byj , (2R0 j0 , )
0

0

=O
By (3R0 j0 , )
dg (y , y)
2n
dvg

= O j20 ,
so that


1 n
= o j0 ,2
(6.2.37)
where
= Byj0 , (2R0 j0 , ) \Byj0 , (R0 j0 , ) .
Similarly, it is easily checked that

G (y , y) h (y) (y) u (y) vNj0 , (y) dvg

1 n
= o j0 ,2 .
(6.2.38)
157
Independently, from (6.0.1), (6.1.80), and the definition of vNj0 , ,

1 n
G (y , y) h (y) vNj0 , (y) dvg = o j0 ,2 .
Byj , (R0 j0 , )
0
(6.2.39)
Coming back to (6.2.36), using (6.2.37)(6.2.39), we then get that

u (y ) vNj0 , (y )
=
G (y , y) (g u (y) + h (y) u (y)) dvg
Byj , (R0 j0 , )
0

1 n
G (y , y) g vNj0 , (y) dvg + o j0 ,2

Byj , (R0 j0 , )
0
2 1
=
dvg
G (y , y) u (y)
Byj , (R0 j0 , )
0

1 n
G (y , y) g vNj0 , (y) dvg + o j0 ,2 .

Byj , (R0 j0 , )
0
Passing through geodesic normal coordinates at yj0 , , it is easily checked that

1 n
2 1
g vNj0 , (y) = vNj0 , (y)
+ O j0 ,2
in Byj0 , (R0 j0 , ). Thus, as above,
G (y , y) g vNj0 , (y) dvg

Byj , (R0 j0 , )
0
2 1
=
dvg
G (y , y) vNj0 , (y)
Byj , (R0 j0 , )
0

1 n
2
+O j0 ,
G (y , y) dvg
Byj , (R0 j0 , )
0

1 n
2 1
=
dvg + o j0 ,2
G (y , y) vNj0 , (y)
Byj , (R0 j0 , )
0
and it follows that
u (y ) vNj0 , (y )

2 1
2 1
=
dvg
vNj0 , (y)
G (y , y) u (y)
Byj , (R0 j0 , )
0

1 n
+ o j0 ,2 .
(6.2.40)
From point (P1) of Theorem 4.1, up to a subsequence, there exists 0 as

+ such that
(
sup
Byj , R0 j0 , \rj=1 Byj,

0
( j0 , )
j20 , |u vNj0 , | 0
(6.2.41)
158
CHAPTER 6
as +. Using (6.1.80), as above, one then gets that

2 1
2 1
dvg
vNj0 , (y)
G (y) u (y)
Byj , (R0 j0 , )\rj=1 Byj, ( j0 , )
0

1 n
2
2n
= o j0 ,

1 n
= o j0 ,2 ,
Byj
(R0 j0 , )
0 ,
dg (y , y)
dvg
where G (y) = G (y , y). Hence, (6.2.40) can be rewritten as

u (y ) vNj0 , (y )

1 n
2 1
dvg + o j0 ,2 .
G (y , y) u (y)
rj=1 Byj, ( j0 , )
(6.2.42)
Given j = 1, . . . , r, we define

(y) = dg (
yj, , y) ,
j, = y M s.t. R

1
1
= y j, Byj, ( j0 , ) s.t. dg (y , y) R
(y
)
,
Bj,
2

1
2
= y j, Byj, ( j0 , ) s.t. dg (y , y) < R
(y
)
.
Bj,
2
1
n+2 .
We fix 0 < <
Let j {1, . . . , r}. We write that

2 1
j, Byj, ( j0 , )
G (y , y) u (y)
dvg
2 1
j, Byj, (R()
1, )
G (y , y) u (y)
2 1
1 \B
Bj,
1, )
y
j, (R()
G (y , y) u (y)
dvg
2 1
2 \B
Bj,
1, )
y
j, (R()
G (y , y) u (y)
(6.2.43)
dvg
dvg
where R () is as in claim 6.2.2. We assume in (6.2.43) that 1, = o ( j0 , ).

This is always possible up to changing . Since
(y )
R
1,
+ as +,
dg (y , y) dg (y , yj, ) dg (
yj, , y)
R (y ) R () 1,
1
R
(y )
2
for large and all y j, Byj, (R () 1, ). Using (6.1.80) we then get that
159
for large
2 1
G (y , y) u (y)
j, Byj, (R()
1, )
(y )
C2 2n2 R
2n
dvg
2 1
u
dvg
j, Byj, (R()
1, )

1

u 22 1 Volg Byj, R () 1, 2

21
n
(y )2n 2 1 2 n1 R ()n 1,
C2 2n2 R
.
n
Hence, for large,
2 1
G (y , y) u (y)
dvg
C2 2
n2
(y )
R
2n
j, Byj, (R()
1, )
n
2 1
(6.2.44)
(y )
D1 1, R
2n
where
D1 = 2n1 C2 2
n1
21
n
1
, we can write using (6.1.80) that
Given y Bj,
R () 2
(y )
G (y , y) C2 2n2 R
2n
(6.2.45)
Independently, it comes from (6.2.4) that Byi, (R () i, ) i, for large and

all i {1, . . . , r}. Applying claim 6.2.2, we get that
! n +1 (12)

(2 )
2 1
2 1
(n+2)(1)
u (y)
1,
22 2 C ()
dg (
yj, , y)
"
( n +1)(12)
(n+2)
+j0 ,2
dg (
yj, , y)
1
\Byj, (R () 1, ). Hence,
for all y Bj,
2 1
G (y , y) u (y)
dvg
1 \B
Bj,
1, )
y
j, (R()
22
C ()
+22
1
n+2 ,
( n +1)(12)
j0 ,2
(n+2)
1 \B
Bj,
1, )
y
j, (R()
Since <
2 1
+n4
2 1
C ()
G (y , y) dg (
yj, , y)
n
(y )2n ( 2 +1)(12)
C2 R
1,
(n+2)(1)
1 \B
Bj,
1, )
y
j, (R()
dg (
yj, , y)
(n+2)(1)
1 \B
Bj,
1, )
y
j, (R()
dg (
yj, , y)
dvg .
dvg
2n1
n(n+2)(1)
(R () 1, )
(n + 2) (1 ) n
dvg
160
CHAPTER 6
for large. Independently, from Holders inequality and (6.1.80),
(n+2)
G (y , y) dg (
yj, , y)
dvg
1 \B
Bj,
1, )
y
j, (R()
= O
n(n2)
n1
By (2R0 j0 , )
dg (y , y)

O

Byj, ( j0 , )
since <
1
n+2 .
n1
n(n+2)
yj, , y)
dvg
dg (
1(n+2)
= O j0 , ( j0 , )
Hence,
1 \B
Bj,
1, )
y
j, (R()
n1
n
dvg
G (y , y) dg (
yj, , y)
(n+2)

2(n+2)
dvg = o j0 ,
since 0 as +. Summarizing,
2 1
G (y , y) u (y)
dvg
1 \B
Bj,
1, )
y
j, (R()
n
2 1
D2 1,
(6.2.46)

n
(y )2n + o 1 2
R
j0 ,
where

D2 = 22 +n3
n1
2 1
n(n+2)(1)
C2 C ()
R ()
. (6.2.47)
(n + 2) (1 ) n
2
In Bj,
, we have that
dg (
yj, , y) dg (
yj, , y ) dg (y , y)
(y )
(y ) 1 R
R
2
1
R
(y ) .
2
Then, applying claim 6.2.2, we obtain as above, using (6.2.3), that
(n+2)(1)
! n +1 (12) 1

(2 )
2 1
2 1
(y )
u (y)
R
1,
22 2 C ()
2
"
n
( +1)(12)
(n+2)
+j0 ,2
dg (
yj, , y)
2
for all y Bj,
\Byj, (R () 1, ). Hence,
2 1
G (y , y) u (y)
dvg
2 \B
Bj,
1, )
y
j, (R()
( n2 +1)(12)
(n+2)(1)
= O 1,
R (y )

+O
( n +1)(12)
j0 ,2
2
Bj,
2
Bj,
G (y , y) dvg
G (y , y) dg (
yj, , y)

(n+2)
dvg
161
As above, we have that

2(n+2)
(n+2)
.
G (y , y) dg (
yj, , y)
dvg = o j0 ,
2
Bj,
By (6.1.80), we also have that

G (y , y) dvg = O
2
Bj,
By
(y )
R
2
2 1
2 \B
Bj,
1, )
y
j, (R()
G (y , y) u (y)
n
2 1
(y )
= O 1, R
(y )
R
1,
dg (y , y)
dvg

(y )2 .
=O R
Therefore,
Since
2n
2n
1,
(y )
R
dvg
2(n+2)

1 n
+ o j0 ,2 .
1
+ as +, and since < n+2
, we obtain that
2 1
G (y , y) u (y)
dvg
2 \B
Bj,
1, )
y
j, (R()
n
2 1
= o 1,

n
(y )2n + o 1 2 .
R
j0 ,
Combining (6.2.43), (6.2.44), (6.2.46), and (6.2.48), we get that
2 1
dvg
G (y , y) u (y)
j, Byj, ( j0 , )

n
1 n
2n
2 1
(D1 + D2 + o (1)) 1,
+ o j0 ,2
R (y )
(6.2.48)
(6.2.49)
for all j {1, . . . , r}. Since

rj=1 Byj, ( j0 , ) = rj=1 j, Byj, ( j0 , ) ,

n
1 n
2n
2 1
+ o j0 ,2 .
R (y )
u (y ) vNj0 , (y ) r (D1 + D2 + o (1)) 1,
This proves that (6.2.33) holds with C > r (D1 + D2 ), where D1 is as in (6.2.45)
and D2 is as in (6.2.47).
Thanks to cases 14, (6.2.33) holds for any sequence (y ) of points in

0 j,
Byj0 , (R0 j0 , ) \ rj=1 Byj, R
if we choose C such that C > D0 + r (D1 + D2 ), where D0 is as in (6.2.34), D1
is as in (6.2.45), and D2 is as in (6.2.47). This ends the proof of claim 6.2.3. 2
From now on, we let j {1, . . . , r}, and let
j, (y) = min dg (
yi, , y) .
R
i=j,...,r
(6.2.50)
162
CHAPTER 6
We say that (6.2.51)j holds if there exist C > 0 and a sequence ( ) of positive
real numbers converging to 0 as + such that
u (y) vNj0 , (y)
j1

n 1
(6.2.51)j
n
n
1
1
2n
2n
2
2
j, (y)
C
i,
dg (
yi, , y)
+ j,
+ j0 ,2
R
i=1

0 i, and all . The objective of
for all y Byj0 , (R0 j0 , ) \ ri=1 Byi, R
claims 6.2.4 and 6.2.5 is to prove that, for any j {2, . . . , r},
(6.2.51)j1 (6.2.51)j .
(6.2.52)
We let j {2, . . . , r} and we assume that (6.2.51)j1 holds. Before stating claim
6.2.4, we prove some preliminary results and set up some notation. We fix such
1
. Given k = 1, . . . , j 1, we let
that 0 < < n+2

k, = y Byj0 , (R0 j0 , ) s.t. k, (y) i, (y) , i = 1, . . . , j 1
(6.2.53)
where
( n 1)(12)
(2n)(1)
i, (y) = i,2
dg (
yi, , y)
.
(6.2.54)
0 such that,
We claim now that there exists R () R
n2 1

j, (y) , dg (
yk, , y)
u (y) vNj0 , (y)
min R
1

min {D () , D (1 )} 2 2
<
100
(6.2.55)
for large, for all k {1, . . . , j 1} and all y k, \ ri=1 Byi, (R () i, ),

where D () is as in (6.2.30). The proof of this claim is based on (6.2.51)j1
i, )
and claim 6.2.1. We let (y ) be a sequence of points in k, \ ri=1 Byi, (R
for some k {1, . . . , j 1}, R R0 . We need to prove that we can choose R

sufficiently large such that

n2 1

j, (y ) , dg (
min R
yk, , y )
u (y ) vNj0 , (y )
1

(6.2.56)
min {D () , D (1 )} 2 2
<
100
for large. Applying (6.2.51)j1 , we get that
u (y ) vNj0 , (y )
j2

n 1
n
n
1
1
2n
2n
2
2
j1, (y )
C
.
i,
dg (
yi, , y )
+ j1,
+ o j0 ,2
R
i=1
Since

j, (y ) , dg (y , yj1, ) ,
j1, (y ) = min R
R
163
we have that
j, (y )2n + dg (y , yj1, )2n .
j1, (y )2n R
R
Hence,
u (y ) vNj0 , (y )
j1

n 1
n
1 n
2n
2n
2
2 1
2
C
.
i, dg (
yi, , y )
+ j1, Rj, (y )
+ o j0 ,
i=1
i, for all i {1, . . . , r},

Since dg (y , yi, ) R
n
2n
2
i,
dg (
yi, , y )
R(2n) i, (y )
for all i {1, . . . , j 1}. Since we have that y k, , we also have that
i, (y ) k, (y ) for all i {1, . . . , j 1}. Thus,
u (y ) vNj0 , (y )
"

!
n
1 n
2n
2 1
,
+ o j0 ,2
C (j 1) R(2n) k, (y ) + j1,
Rj, (y )
yk, , y ) = O (j0 , ), we have that
and since dg (
n2 1

j, (y ) , dg (
yk, , y )
u (y ) vNj0 , (y )
min R

n2
2
!
"
n2
j1,
(2n)
2
C (j 1) R
dg (
yk, , y )
k, (y ) +
+ o (1) .
j, (y )
R
yk, , y ) R
k, , we get that
As a consequence, noting that dg (
n2 1

j, (y ) , dg (
yk, , y )
u (y ) vNj0 , (y )
min R
n2 1

n
j1,
+ o (1) .
C (j 1) R1 2 +
j, (y )
R
1
j, (y ) + as +, it is clear that (6.2.56) holds up to
If j1,
R
j, (y ) = O (
choosing R sufficiently large. Let us assume now that R
j1, ).
We claim that this implies that Rj, (y ) = R (y ) for large. In order to prove
j, (y ) = dg (y , yl, ). For any
this claim, we let l {j, . . . , r} be such that R
i {1, . . . , j 1}, from (6.2.3),
yl, , yi, ) dg (y , yl, )
dg (y , yi, ) dg (
j1,
j1,
j1,
j, (y )
yl, , yi, ) i,
dg (
R
i,
j1,
j1,
yl, , yi, )
dg (
O (1) .
i,
By (6.2.4) we then get that

dg (y , yi, )
+
j1,
164
CHAPTER 6
as +, and it follows that for any i with i {1, . . . , j 1}, we have that
j, (y ) when is large. In particular, if R
j, (y ) = O (
yi, , y ) > R
j1, ),
dg (
then we have Rj, (y ) = R (y ) for large, and this proves the above claim.
Thanks to claim 6.2.1 we then get that (6.2.56) holds up to choosing R sufficiently
large. In particular, (6.2.55) is proved.
We let now Aj, be defined by
( n2 1)(12)
Aj,
=
where
sup
k, (y)
sup
k=1,...,j1 yk,
( n 1)(12)
(2n)(1)
j,2
Rj, (y)
(6.2.57)
k,
= y k, \ ri=1 Byi, (R () i, ) s.t.

n
yk, , y) 2
dg (

u (y) vNj0 , (y)
D (1 )
100
212
.
(6.2.58)
By convention, Aj, = if the

possibly larger R (),
lim sup
+
k, s
are empty. We claim that for some
Aj, j,
1.
j1,
(6.2.59)
We prove this claim in what follows. If Aj, = then, up to a subsequence,

there exist k {1, . . . , j 1} and a sequence (y ) of points in k, such that
k, (y ) = (Aj, j, )( 2
1)(12)
j, (y )(2n)(1) (1 + o (1)) .
R
We assume by contradiction that (6.2.59) is false. Then, up to a subsequence,

( n2 1)(12)
(2n)(1)
.
Rj, (y )
k, (y ) j1,
(6.2.60)
k, ,
Since y
1

n
D (1 ) 2 2
1
dg (
yk, , y ) 2
u (y ) vNj0 , (y ) .
100
We then get, thanks to (6.2.51)j1 , that
j1

1
n 1
n
D (1 ) 2 2
1
2n
2
Cdg (
yk, , y ) 2
i,
dg (
yi, , y )
100
i=1

n
2n
2 1
+
+ o (1) .
Rj, (y )
j1,
yi, , y ) R () i, for all i {1, . . . , j 1} and all .

Since y k, , dg (
Moreover, k, (y ) i, (y ) for all i {1, . . . , j 1}. Hence,
j1

i=1
2n
2
i,
dg (
yi, , y )
(2n)
(j 1) R ()
k, (y ) ,
165
and we get that

1

!
n
D (1 ) 2 2
1
(2n)
Cdg (
yk, , y ) 2
k, (y )
(j 1) R ()
100
"
n
2n
2 1
+ o (1) .
+
j1,
Rj, (y )
Using (6.2.60), we can write that

(n2)
n
j1,
( n2 1)(12)
2n
(2n)(1)
2 1
j1, Rj, (y )
=
j1,
Rj, (y )
Rj, (y )

(n2)
j1,
k, (y ) .
j, (y )
R
Since dg (
yk, , y ) R () k, , we also have that
n
yk, , y ) 2
dg (
k, (y ) R ()(
1 n
2 )(12)
Combining the three equations above, it follows that

(n2)
1

n
D (1 ) 2 2
j1,
1
CR () 2 (j 1) + R ()
o (1) .
j, (y )
100
R
Taking R () sufficiently large, this gives that
j, (y ) = O (
R
j1, ) .
(6.2.61)
Coming back to (6.2.60), we get that

n 1 (12)

k, ( 2 )
(n2)(1)
j, (y )(n2)(1)
dg (
yk, , y )
R
j1,

n
( 2 1)(12) n2 1
= O k,
j1, .
Since k j 1, we have by (6.2.3) that j1, k, . Thus,
dg (
yk, , y ) = O (
k, ) .
(6.2.62)
j, (y ) = dg (
yl, , y ). Thanks to
Now we let l {j, . . . , r} be such that R
(6.2.61) and (6.2.62) we can write that
dg (
yk, , yl, ) dg (
yk, , y ) + dg (
yl, , y )
j1, )
= O (
k, ) + O (
= O (
k, ) .
Noting that such an equation is in contradiction with (6.2.4), we get that (6.2.59)
is proved.
We let
j, = j, max {Aj, ; 1}
(6.2.63)
166
CHAPTER 6
and prove the following:

C LAIM 6.2.4. Let j {2, . . . , r} and assume that (6.2.51)j1 holds. Let also
1
0 be such that (6.2.55) and (6.2.59)
, and R () R
be such that 0 < < n+2
hold. Then there exists C () > 0 such that
j1

( n 1)(12)
(2n)(1)
i, (y) + 2
u (y) C ()
Rj, (y )
j,
i=1
(1 n )(12)
(2n)
+j0 , 2
R (y)
(y) is
for all y Byj0 , (R0 j0 , ) \ ri=1 Byi, (R () i, ) and all , where R
as in claim 6.2.1, Rj, (y) is as in (6.2.50), and j, is as in (6.2.63).

Proof of claim 6.2.4. We let x0 SNj0 , where SNj0 is as in Theorem 4.1, and we
let x = expyj0 , (j0 , x0 ). It is a direct consequence of (6.2.55) that there exists

0 < () 12 d x0 , SNj0 \ {x0 } such that
(y)2 u (y)2
R
D ()
(6.2.64)
for all y Bx ( () j0 , ) \ ri=1 Byi, (R () i, ) and large. We consider the

linear operator L given by
2 2
u.
L (u) = g u + h u u
We let
H (y) =
j1

( n 1)(12)
1
i,2
Gi, (y)
i=1
( n 1)(12)
1
+j,2
Gi, (y)
r
(1 n )(12)
+j0 , 2
i=j
r
Gi, (y)
i=1
where
yi, , y) ,
Gi, (y) = G (
and G is the Greens function of g +h0 , where h0 is as in the beginning of section
6.1. We have already seen when proving claim 6.2.2 that
r

(y)(2n)2
L
Gi, (y) 2rD () C2 R
(6.2.65)
i=1
for all y Bx ( () j0 , ) \ ri=1 Byi, (R () i, ) and large. We just need to
167
replace (6.2.31) by (6.2.64). Then, easy computations give that

L H (y)
j1

( n 1)(12)
1
i,2
Gi, (y)
i=1
(1 )
|Gi, (y) |2
2
Gi, (y)

2 2
u (y)
+ h (y) (1 ) h0 (y)
( n 1)(12)
1
+ j,2
Gi, (y)
r
i=j

(1 )
|Gi, (y) |2
2
Gi, (y)

2 2
u (y)
+ h (y) (1 ) h0 (y)
(1 n )(12)
(2n)2
+ 2rD () C2 j0 , 2
R (y)
(6.2.66)
ri=1
Byi, (R () i, ) and large. Let (y ) be a

for all y Bx ( () j0 , ) \
sequence of points in Bx ( () j0 , ) \ ri=1 Byi, (R () i, ). We claim that
L H (y ) 0
for large. By (6.1.2), we have that, for large,
|Gi, (y ) |2
2
Gi, (y )
C12
dg (
yi, , y )
for all i {1, . . . , r}. Thus, thanks to (6.2.66), we can write that, for large,
L H (y )
j1

( n 1)(12)
1
i,2
Gi, (y )
i=1
C12
(1 )
2
dg (
yi, , y )
2 2
u (y )
+ O (1)
( n 1)(12)
1
+ j,2
Gi, (y )
r
i=j
(6.2.67)
C12
(1 )
2
dg (
yi, , y )
2 2
u (y )
+ O (1)
(1 n )(12)
(2n)2
+ 2rD () C2 j0 , 2
.
R (y )
After passing to a subsequence, we assume that y k, for some 1 k j 1.
We distinguish four cases.

j, (y ) , dg (
yk, , y )
min R
j0 ,
168
CHAPTER 6
as +, for some > 0. Then, by (6.2.55),

2 2
= O j2
.
u (y )
0 ,

2 2
u (y )
+ O (1)
j2
0 ,
j1
i,
i=1
2
O j0 , k,
( n 1)(12)
1
i,2
Gi, (y )
i=1

=O
j1

(y )
(y )
since y k, . By (4.1.6) and the construction of the yj, s, k, = o (j0 , ).

We also have here that for large, dg (
yk, , y ) 2 j0 , . Hence,

2 2
u (y )
=o
+ O (1)
1 n
j0 , 2
j1

( n 1)(12)
1
i,2
Gi, (y )
i=1
(6.2.68)
Similarly, thanks in particular to (6.2.59), we have that

2 2
u (y )
=o
r
n 1 (12)
)
(
1
+ O (1) j,2
Gi, (y )
1 n
j0 , 2
(6.2.69)
i=j

.
Combining (6.2.67) with (6.2.68) and (6.2.69), we then get that, for large,

(1 n )(12)
1 n
(2n)2
L H (y ) 2rD () C2 j0 , 2
+ o j0 , 2 .
R (y )
(y ) 2R0 j , , it follows from this equation that L H (y ) 0
Since R
0
when is large.
j, (y ), and that
yk, , y ) R
Case 2. We assume that y k, , that dg (
j, (y ) = o (j , ). By (6.2.55), we then get that
R
0
j, (y )2 u (y )2
R
D (1 )
for large, and it follows [see, for instance, the proof of (6.2.32)] that
( n 1)(12)
1
Gi, (y )
j,2
r
i=j
C12
(1 )
2
dg (
yi, , y )
2 2
u (y )
+ O (1)
( n 1)(12)
(2n)(1)2
2rD (1 ) C21 j,2
Rj, (y )
(6.2.70)
169
for large. Since y k, ,

n
yk, , y ) 2
dg (

u (y ) vNj0 , (y )
D (1 )
100
212
so that
2
2 2
dg (
yk, , y ) u (y )

max 1, 22 3

D (1 )
2
2 2
+ dg (
.
yk, , y ) vNj0 , (y )
100
(6.2.71)
1 n2
d (
y
,y )
If g k,
as
+
for
some
>
0,
then
u
(y
)
=
O
j0 , . It
j0 ,
follows, as in case 1, that
j1

( n 1)(12)
2 2
1
+ O (1)
i,2
Gi, (y )
u (y )
(6.2.72)
i=1

1 n
2
= o j0 ,
.
yk, , y ) = o (j0 , ), we get with (6.2.71) that
If dg (
2
2 2
dg (
yk, , y ) u (y )
D (1 )
for large. Then, from (6.1.3),

j1

( n2 1)(12)
1
Gi, (y )
i,
C12
i=1
(1 )
2 2
u (y )
dg (
yi, , y )
+ O (1)
(1 )
( n2 1)(12) C12
(2n)(1)
k,
dg (
yk, , y )
2
C21
dg (
yk, , y )

2
C21 D (1 ) dg (
yk, , y ) + O (1)
j1

( n 1)(12)
(2n)(1)
dg (
yi, , y )
i,2
i=1
(1 )
C12
2
dg (
yk, , y ) k, (y )
C21
j1

2
yk, , y ) + O (1)
i, (y ) .
C21 D (1 ) dg (
i=1
Since y k, , and from the definition (6.2.30) of D (1 ), this gives that

j1

( n 1)(12)
1
i,2
Gi, (y )
i=1
C12
(1 )
2
dg (
yi, , y )
2 2
u (y )
(6.2.73)
+ O (1)
2
yk, , y )
(4r (j 1) + o (1)) C21 D (1 ) dg (
k, (y ) .
170
CHAPTER 6
Combining (6.2.67) with (6.2.70), (6.2.72), and (6.2.73), we then get, as in case
1, that L H (y ) 0 for large.
j, (y ) and that
Case 3. We assume that y k, , that dg (
yk, , y ) R
dg (
yk, , y ) = o (j0 , ). We then get with (6.2.55) that
2
2 2
yk, , y ) u (y )
dg (
D (1 )
for large. As when proving (6.2.73), it follows that

j1

( n 1)(12)
1
Gi, (y )
i,2
i=1
C12
(1 )
2
dg (
yi, , y )
2 2
u (y )
(6.2.74)
+ O (1)
2
yk, , y )
(4r (j 1) + o (1)) C21 D (1 ) dg (
k, (y ) .
j, (y ) dg (
yk, , y ), and since
Since R
2 2
u (y )
D (1 ) dg (
yk, , y )
we can write with (6.1.3) that

r

( n 1)(12)
2 2
1
j,2
+ O (1)
Gi, (y )
u (y )
i=j
(r j + 1) D (1 ) dg (
yk, , y )

+ O (1) C21
( n 1)(12)
(2n)(1)
j,2
Rj, (y )
( n 1)(12)
2+(2n)(1)
(r j + 1 + o (1)) D (1 ) C21 j,2
dg (
yk, , y )
.
By (6.2.3), (6.2.59), and (6.2.63),
j, k, (1 + o (1))
since k j 1. It follows that
r

( n 1)(12)
2 2
1
j,2
+ O (1)
Gi, (y )
u (y )
(6.2.75)
i=j
(r j + 1 + o (1)) D (1
) C21 dg
(
yk, , y )
k,
(y ) .
Combining (6.2.67) with (6.2.74) and (6.2.75), we then get that L H (y ) 0

for large.
Case 4. We assume that y k, and that

j, (y ) ; dg (
yk, , y ) = o (j0 , ) .
min R
Since y k, ,
dg (
yk, , y )
n
2 1
u (y ) vNj0 , (y )
D (1 )
100
212
.
171
By (6.2.55) we then get that

j, (y ) dg (
yk, , y ) ,
R
j, (y ) = o (j , ), we have that
and since R
0
j, (y )2 u (y )2
R
D (1 )
for large. As when proving (6.2.32), it follows that

r
( n 1)(12)
1
j,2
Gi, (y )
i=j
C12
(1 )
2
dg (
yi, , y )
2 2
u (y )
+ O (1)
(6.2.76)
(3r + j 1 + o (1)) D (1 ) C21

( n 1)(12)
(2n)(1)2
.
j,2
Rj, (y )
On the other hand, from (6.1.3) and the above equations,
j1

( n 1)(12)
2 2
1
u (y )
+ O (1)
i,2
Gi, (y )
i=1
j1

j, (y )2 + O (1) C 1
D (1 ) R
i, (y )
2
i=1
(j 1 +
o (1)) C21 D (1
j, (y )2 (y )
) R
k,
since y k, . Noting that y also belongs to k, , we get from the definitions

(6.2.57) of Aj, and (6.2.63) of j, that
n
1 (12)
(2n)(1)
(y ) (A )( 2 )
R (y )
k,
j, j,
(
j,
n
2 1
j,
)(12)
(2n)(1)
.
Rj, (y )
This leads to
j1

( n 1)(12)
2 2
1
+ O (1)
i,2
Gi, (y )
u (y )
i=1
(j 1 + o (1)) C21 D (1 )
(6.2.77)
( n 1)(12)
(2n)(1)2
j,2
.
Rj, (y )
Combining (6.2.67) with (6.2.76) and (6.2.77), we then get that L H (y ) 0
for large.
The study of these four cases clearly proves that, for large,
L H 0
(6.2.78)
ri=1
in Bx ( () j0 , ) \
Byi, (R () i, ). Using point
4.1,

(P1) of Theorem
(6.1.3), (6.2.3), (6.2.63), and the fact that () 12 d x0 , SNj0 \ {x0 } , we also
have that there exists C () > 0 such that
u C () H
(6.2.79)
172
CHAPTER 6
on Bx ( () j0 , ) \ ri=1 Byi, (R () i, ) . By [9], L satisfies the maximum principle. Since L u = 0, we then get with (6.2.78), (6.2.79), and (6.1.3)
that the equation of claim 6.2.4 holds in

x0 SNj Bx ( () j0 , ) \ ri=1 Byi, (R () i, ) .
0
By point (P1) of Theorem 4.1, the equation extends to

Byj0 , (R0 j0 , ) \ ri=1 Byi, (R () i, ) .
This proves claim 6.2.4.
Now that claim 6.2.4 is proved, we prove that (6.2.52) holds. This is the subject
of claim 6.2.5. Claim 6.2.5 is as follows:
C LAIM 6.2.5. If (6.2.51)j1 holds for some j = 2, . . . , r, then (6.2.51)j holds
also.
Proof of claim 6.2.5. We claim that there exists C > 0 such
that for any sequence
r
(y ) of points in Byj0 , (R0 j0 , ) \ i=1 Byi, R0 i, ,

u (y ) vNj0 , (y )
j1

n
1 n
2n
0
2 1
C
+ o j0 ,2
i, (y ) + j, Rj, (y )
(6.2.80)
i=1
where
n
2n
2
dg (
yi, , y)
0i, (y) = i,
(6.2.81)
1
in the definition (6.2.57)
and j, is as in (6.2.63). We fix such that 0 < < n+2
of Aj, . Before proving (6.2.80), we introduce some notation and discuss some
consequences of claim 6.2.4. Given i = 1, . . . , j 1, we let

( n2 1)(12)
(2n)(1)
(6.2.82)
Rj, (y)
i, = y i, s.t. i, (y) j,
where i, is as in (6.2.53), and i, is as in (6.2.54). We let also

= By
.
(R0 j , ) \ j1
j,
j0 ,
i=1
i,
(6.2.83)
As in (6.1.127) and (6.1.129), for large,

, and
if k {1, . . . , j 1} then Byk, (R () k, )
k,
if k {j, . . . , r} then By (R () j, ) .
k,
(6.2.84)
j,
By claim 6.2.4, (6.2.3), (6.2.63), and (6.2.84) [see also the proof of (6.1.128)] we
\By (R () k, ),
get that for any k {1, . . . , j 1}, and any y
k,
k,
$
#
n

( +1)(12)
2 1
2 1
(y)(n+2)
u (y)
(6.2.85)
C0 k, (y)
+ j0 ,2
R
when is large, and where C0 does not depend on . By claim 6.2.4, (6.2.3),
\r By (R () j, ),
(6.2.63), and (6.2.84), we also have that, for any y
i,
j,
i=j
!
n
( 2 +1)(12)
2 1
(n+2)(1)
C0 j,
u (y)
Rj, (y)
(6.2.86)
"
n
( 2 +1)(12)
(n+2)
+ j0 ,
R (y)
173
when is large, and where C0 is as above. We let ( ) be a sequence of positive

real numbers converging to 0 as + such that
(
sup
Byj , R0 j0 , \ri=1 Byi,

0
( j0 , )
j20 , |u vNj0 , | 0
(6.2.87)
as +. The existence of ( ) easily follows from point (P1) of Theorem 4.1.

Now we prove that (6.2.80) holds. For that purpose, we let (y ) be such that

0 i,
y Byj0 , (R0 j0 , ) \ ri=1 Byi, R
for all , and distinguish four cases.
point (P1) of Theorem 4.1,
(y )
R
j0 ,
as + for some > 0. Then, from

1 n
|u (y ) vNj0 , (y ) | = o j0 ,2 .

0 such that
Case 2. We assume that there exists k {1, . . . , r} and R R
(
yk, , y ) R as +. Then, as in case 2 of the proof of claim 6.2.4,
1
dg
k,
lim sup
+
u (y )
D0
0k, (y )
where
D0 = sup
sup |z|n2 u (z yk )
i=1,...,r zRn
(6.2.88)
and yk = xNj0 +qk is as in Theorem 4.1. By (6.2.3) and (6.2.63) we then get that
(6.2.80) holds with C > D0 .
1
dg (
yk, , y ) + as + for all k such
Case 3. We assume that k,
j, (y ) = O (j, ). Then, as in case 3 of the proof
that k {1, . . . , r}, and that R
(y ). Since
j, (y ) = R
of claim 6.1.8, we get that for large, R
1
dg (
yk, , y ) +
k,
as + for all k {1, . . . , r}, claim 6.2.1 gives that

(y ) 2 1 |u (y ) vN , (y ) | = o (1) .
R
j0
n
Hence,

1 n
n2
j, 2 R
u (y ) vNj0 , (y )
j, (y )
n 1

j, (y ) 2

n
R
(y ) 2 1 u (y ) vN , (y )
R
=
j0
j,

n2 1
Rj, (y )
= o
j,
= o (1) .
174
CHAPTER 6

(y ) = o (j , ), that 1 dg (
yk, , y ) + as
Case 4. We assume that R
0
k,
1
+ for all k {1, . . . , r}, and that j, Rj, (y ) + as the parameter
+. We let ( ) be a sequence of smooth functions with compact support in
Byj0 , (2R0 j0 , ) such that 1 in Byj0 , (R0 j0 , ) and

, g = O j2
.
= O j1
0 ,
0 ,
We let G be the Greens function of g +h . Thanks to the Greens representation
formula,
u (y ) vNj0 , (y )

=
Byj , (2R0 j0 , )
0

+
G (y , y) h (y) (y) u (y) vNj0 , (y) dvg .
Byj , (2R0 j0 , )
0
As in case 4 in the proof of claim 6.2.3, this equation reduces to

1 n
u (y ) vNj0 , (y ) = o j0 ,2

2 1
2 1
+
dvg .
vNj0 , (y)
G (y , y) u (y)
Byj , (R0 j0 , )
0
Using (6.1.80) and (6.2.87), we then get that
u (y ) vNj0 , (y )

1 n
2 1
dvg + o j0 ,2
G (y , y) u (y)
ri=1 Byi, ( j0 , )
(6.2.89)
where 0 is as in (6.2.87). Given k {1, . . . , j 1}, we let

r

1
1
yk, , y ) ,
Bk, = y k, i=1 Byi, ( j0 , ) s.t. dg (y , y) dg (
2

r

1
2
yk, , y ) .
Bk, = y k, i=1 Byi, ( j0 , ) s.t. dg (y , y) < dg (
2
Then,
2 1
k,
ri=1 Byi,
2 1
1
Bk,
G (y , y) u (y)
+
Since
dg (
yk, ,y )
k,
( j0 , ))
G (y , y) u (y)
2
Bk,
dvg
2 1
G (y , y) u (y)
dvg
(6.2.90)
dvg .
+ as +, we may assume that

1
Byk, (R () k, ) Bk,
(6.2.91)
175

1
, it follows from (6.1.80) that
for large, up to changing . Given y Bk,
2n
G (y , y) C2 2n2 dg (
yk, , y )
Thus we can write that
2 1
G (y , y) u (y)
dvg
Byk, (R()
k, )
C2 2
n2
2n
dg (
yk, , y )
2 1
Byk, (R()
k, )
2n
C2 2n2 dg (
yk, , y )
2 1
G (y , y) u (y)
where
D1 = 2n2 C2 2
From (6.2.85),
2 1
u (y)
#
C0
k,
dvg
1

u 22 1 Volg Byk, (R () k, ) 2 .
This gives that
Byk, (R()
k, )
u (y)
dvg (D1 + o (1)) 0k, (y ) (6.2.92)
n1
21
2 1
(y)
R () 2
( n +1)(12)
j0 ,2
R
(6.2.93)
(n+2)
(y)
1
for all y Bk,
\Byk, (R () k, ). Noting that
n+2
2
k,
(12)
2n
dg (
yk, , y )
2(n+2)
= 0k, (y ) k,
we can then write with (6.1.80) that
2 1
G (y , y) u (y)
dvg
1 \B
Bk,
k, )
y
k, (R()
( n +1)(12)
C0 j0 ,2
1 \B
Bk,
k, )
y
k, (R()
(y)(n+2) dvg
G (y , y) R
2(n+2)
+C0 C2 2n2 0k, (y ) k,
(n+2)(1)
dg (
yk, , y)
dvg .
1 \B
Bk,
k, )
y
k, (R()
Since <
1
n+2 ,
1 \B
Bk,
k, )
y
k, (R()
dg (
yk, , y)
(n+2)(1)
dvg
n1 + o (1)
n(n+2)(1)
(R () k, )
.
(n + 2) (1 ) n
Given i = 1, . . . , r, let
Ii, = {j = 1, . . . , r s.t. dg (
yi, , yj, ) 2 j0 , } .
176
CHAPTER 6
Noting that in Byi, ( j0 , ),

(y) = min dg (
yj, , y) ,
R
jIi,
and that for j Ii, , Byi, ( j0 , ) Byj, (3 j0 , ), we can write that
(y)(n+2) dvg
G (y , y) R
1 \B
Bk,
k, )
y
k, (R()
Byj
0 ,
n1
n
(R0 j0 , )
G (y , y)
ri=1 Byi, ( j0 , )
n
n1
(y)n(n+2) dvg
R
n1

1(n+2)
= O (j0 , ) O ( j0 , )
.
In particular,
(y)
G (y , y) R
(n+2)
1 \B
Bk,
k, )
y
k, (R()

2(n+2)
dvg = o j0 ,
since 0 as +, and we have proved that
2 1
G (y , y) u (y)
dvg
1 \B
Bk,
k, )
y
k, (R()
(D2 +
o (1)) 0k,
(y ) + o
1 n
j0 ,2
(6.2.94)
where
n1
n(n+2)(1)
R ()
.
(n + 2) (1 ) n
Combining (6.2.92) and (6.2.94), we then get that
2 1
G (y , y) u (y)
dvg
D2 = C0 C2 2n2
1
Bk,
o (1)) 0k,
(D1 + D2 +
(y ) + o
(6.2.95)
(6.2.96)
1 n
j0 ,2
1
2 dg
2
Independently, it is easily checked that dg (
yk, , y)
(
yk, , y ) in Bk,
. Using (6.2.85) and (6.2.91) we can then write that

( n
2 1
2 1
(n+2)
2 +1)(12)
(n+2)(1)
C0 2
k, (y )
+ j0 ,
u (y)
R (y)
2
. Thus,
for all y Bk,
2 1
G (y , y) u (y)
dvg
2
Bk,
=O
( n +1)(12)
j0 ,2

+O
k,
2 1
(y )
(y)
G (y , y) R
(n+2)
2
Bk,
2
Bk,
G (y , y) dvg
dvg
177
As above,
(y)
G (y , y) R
(n+2)
2
Bk,

2(n+2)
.
dvg = o j0 ,
Independently, using (6.1.80), we can write that

2n
G (y , y) dvg = O
dg (y , y)
dvg
2
Bk,
2
Bk,
=O

By
dg y
k, ,y
2
yk, , y )
= O dg (
Thus,
2 1
2
Bk,
G (y , y) u (y)

= O 0k, (y )
Since <
1
n+2

) dg (y , y)
2n
dvg

.
dvg
k,
dg (
yk, , y )
2(n+2)

1 n
+ o j0 ,2 .
and
yk, , y )
dg (
+
k,
as +, it follows that

1 n
2 1
G (y , y) u (y)
dvg = o 0k, (y ) + o j0 ,2 .
2
Bk,
(6.2.97)
Combining (6.2.90), (6.2.96), and (6.2.97), we then get that for any k such that
k = 1, . . . , j 1,
2 1
dvg
G (y , y) u (y)
(r By ( j , ))
i=1
k,
0
i,
(6.2.98)

1 n
(D1 + D2 + o (1)) 0k, (y ) + o j0 ,2
where D1 is as in (6.2.93), and D2 is as in (6.2.95). Given k {j, . . . , r}, we now
let

\ r By (R () j, ) s.t. R
j, (y) = dg (
k, = y
yk, , y) ,
j,
i=j
i,

r

1
1
Bk, = y k, i=1 Byi, ( j0 , ) s.t. dg (y , y) Rj, (y ) ,
2

1
2
r
Bk, = y k, i=1 Byi, ( j0 , ) s.t. dg (y , y) < Rj, (y ) .
2
178
CHAPTER 6
We write that
2 1
(r By ( j , ))
j,
i=1
0
i,
2 1
G (y , y) u (y)
1
Bk,
k=j

k=j
G (y , y) u (y)
dvg
dvg
(6.2.99)
2 1
2
Bk,
G (y , y) u (y)
dvg
2 1
ri=j Byi, (R()j, )
G (y , y) u (y)
dvg .
1
j, (y )
Rj, (y ) + as +, we have that dg (y , y) 12 R
Since j,
for all y ri=j Byi, (R () j, ) when is large. Using (6.1.80) we then get that
2 1
G (y , y) u (y)
dvg
ri=j Byi, (R()j, )
C2 2
n2
j, (y )2n
R
j, (y )
C2 2n2 R
2n
2 1
ri=j Byi, (R()j, )
u (y)
dvg

21

u 22 1 Volg ri=j Byi, (R () j, )
2n
2 1
(r j + 1 + o (1)) D1 j,
Rj, (y )
n
where D1 is as in (6.2.93). Coming back to (6.2.99), it follows that
2 1
dvg
G (y , y) u (y)
j, (i=1 Byi, ( j0 , ))
r

2 1
G (y , y) u (y)
dvg
1
Bk,
k=j
(6.2.100)

k=j
2 1
2
Bk,
G (y , y) u (y)
dvg
2n
2 1
+ (r j + 1 + o (1)) D1 j,
.
Rj, (y )
n
We let k {j, . . . , r}. From (6.1.80),

j, (y )
G (y , y) C2 2n2 R
2n
1
in Bk,
. Using (6.2.86) we also have that
2 1
u (y)
! n +1 (12)
)
(
(n+2)(1)
C0 j,2
dg (
yk, , y)
"
( n +1)(12)
(y)(n+2)
+j0 ,2
R
179

1
. Thus,
in Bk,
2 1
G (y , y) u (y)
dvg
1
Bk,
C0 C2 2
n2
n
j, (y )2n ( 2 +1)(12)
R
j,
( n +1)(12)
+C0 j0 ,2
As above,
1
Bk,
(n+2)(1)
1
Bk,
dg (
yk, , y)
dvg
(y)(n+2) dvg .
G (y , y) R

(y)(n+2) dvg = o 2(n+2) , and
G (y , y) R
j0 ,
(n+2)(1)
dg (
yk, , y)
dvg

1
Bk,
1
Bk,
n1 + o (1)
n(n+2)(1)
(R () j, )
.
(n + 2) (1 ) n
It follows that
2 1
1
Bk,
G (y , y) u (y)
n
2 1
(D2 + o (1)) j,
dvg
(6.2.101)

n
j, (y )2n + o 1 2
R
j0 ,
2
where D2 is as in (6.2.95). Independently, we have in Bk,
that
1
j, (y) = dg (
yk, , y) R
R
j, (y ) .
2
(n+2)(1)

! n +1 (12) R
j, (y )
(2 )
2 1
u (y)
C0 j,
2
"
( n +1)(12)
(y)(n+2)
+j0 ,2
R
2
. Independently, from (6.1.80),
in Bk,

j, (y )2
G (y , y) dvg = O R
2
Bk,
so that
2 1
2
Bk,
G (y , y) u (y)
dvg
( n +1)(12)
(n+2)(1)
= O j,2
Rj, (y )

+O
( n +1)(12)
j0 ,2
2
Bk,
j, (y )2n
= O j, R
G (y , y) dvg
2
Bk,
(y)
G (y , y) R
(n+2)
n
2 1
j, (y )
R
j,
dvg
(n+2)2

n
+ o j1,2 .
0
180
CHAPTER 6
Since <
2
Bk,
1
and j,
Rj, (y ) + as +, it follows that
n

1 n
2 1
2n
2 1
G (y , y) u (y)
dvg = o j,
+ o j0 ,2 .
Rj, (y )
1
n+2
Coming back to (6.2.100), we then get with (6.2.101) and the above equation that
2 1
dvg
G (y , y) u (y)
(r By ( j , ))
j,
i=1
0
i,
n
2n
(6.2.102)
2 1
(r j + 1 + o (1)) (D1 + D2 ) j,
Rj, (y )

1 n
+ o j0 ,2 .
, it follows from (6.2.89), (6.2.98), and
Noting that Byj0 , (R0 j0 , ) = jk=1
k,
(6.2.102) that
u (y ) vNj0 , (y ) (D1 + D2 + o (1))

1 n
0k, (y ) + o j0 ,2
j1

k=1
2n
2 1
+ (r j + 1) (D1 + D2 + o (1)) j,
Rj, (y )
n
where D1 is as in (6.2.93) and D2 is as in (6.2.95). Hence, (6.2.80) holds with,

for instance, C > r (D1 + D2 ).
From cases 14, equation (6.2.80) holds for any sequence (y ) of points in
Byj0 , (R0 j0 , ) \ ri=1 Byi, (R0 i, ) when C > D0 + r (D1 + D2 ), where D0
is as in (6.2.88), D1 is as in (6.2.93), and D2 is as in (6.2.95). In order to prove
that (6.2.51)j holds, and thus to end the proof of claim 6.2.5, it remains to prove
that
j, = O (
j, ) .
(6.2.103)
By (6.2.63), this reduces to proving that Aj, = O (1) where Aj, is as in (6.2.57).
We proceed by contradiction and assume that
Aj, +
(6.2.104)
as +. By the definition of Aj, , there exists k {1, . . . , j 1} and

y k, such that, up to a subsequence,
(Aj, j, )( 2
1)(12)
j, (y )(2n)(1) = k, (y ) (1 + o (1)) .
R
It follows from (6.2.104) that j, = Aj, j, . Hence,

( n 1)(12)
(2n)(1)
(1 + o (1)) .
k, (y ) = j,2
Rj, (y )
Since y

k, ,
k,
D (1 )
100
(y )
212
i,
(6.2.105)
(y ) for all i {1, . . . , j 1}, and

n
dg (y , yk, ) 2

u (y ) vNj0 , (y ) .
181
Using (6.2.80) and (6.2.105) we then get that

1
D (1 ) 2 2
100
j1

n
n
1
1
2n
0
2
j, (y )
Cdg (
yk, , y ) 2
i, (y ) + j, R

+o
i=1
dg (
yk, , y )
j0 ,
Cdg (
yk, , y )

+
n2 1
j, (y )
R
j,
n
2 1
(n2)
R ()
(n2)
j1
i, (y )
i=1
"
( n 1)(12)
(2n)(1)
j,2
+ o (1)
Rj, (y )
!
n
1
(n2)
yk, , y ) 2 k, (y ) (j 1) R ()
Cdg (

(n2)
"
j, (y )
R
+ o (1) .
+
j,
In particular,

1
D (1 ) 2 2
100

1 n (12)
dg (
yk, , y ) ( 2 )
C
K + o (1)
k,
where

K =
d (
y
j, (y )
R
j,
(6.2.106)
(n2)
+ O (1) .
,y )
+ as +. Let us assume by contradicNow we claim that g k,

k,
tion that dg (
yk, , y ) = O (
k, ). Then, coming back to (6.2.105),

n 1 (12)
j, ( 2 )
(n2)(1)
(n2)(1)
=
dg (
yk, , y )
Rj, (y )
k,

n
( 2 1)(12) n2 1
.
k,
= O j,
j1, ). By (6.2.3), j1, k, since k j 1. Thus,

By (6.2.59), j, = O (
j, (y ) = dg (
j, (y ) = O (
k, ). Let l {j, . . . , r} be such that R
yl, , y ).
R
Then,
yk, , yl, ) dg (
yk, , y ) + dg (y , yl, )
dg (
C k,
182
CHAPTER 6
and this is in contradiction with (6.2.4). In particular, the above claim is proved and
dg (
yk, ,y )
j, (y ) = o (j, ),
+ as +. Then, thanks to (6.2.106), R
k,
and coming back to (6.2.105) we get that

n 1 (12)
k, ( 2 )
(n2)(1)
j, (y )(n2)(1)
dg (
yk, , y )
=
R
j,
n

( 2 1)(12) n2 1
= o k,
j,

n
( 2 1)(12) n2 1
j1,
= o k,

(n2)(1)
.
= o k,
Since
dg (
yk, , y )
+
k,
as +, this is the contradiction we were looking for. In particular, (6.2.104)
is absurd, and (6.2.103) is proved. As already mentioned, this ends the proof of
claim 6.2.5.
2
we then get that (6.1.51)j is true for all j = 1, . . . , r. In particular, (6.2.51)r is
true. Noting that (6.2.51)r implies (6.2.6), we have proved that (6.2.6) holds. As
already mentioned, by finite induction, repeating the above arguments, we in fact
proved that (6.2.2) is true. Then, (6.2.1) is also true, and the upper estimate on the
u s of Theorem 6.1 holds.
6.3 ASYMPTOTIC BEHAVIOR
We prove the estimate from below in Theorem 6.1. For that purpose, we let (x ) be
a sequence of points in M and compute an exact asymptotic expansion of u (x ).
We split this section into two subsections. The first subsection is concerned with
the case u0 0. The second subsection is concerned with the case u0 0.
First we introduce some notations. After passing to a subsequence, we can assume that (x ) converges and we let
x = lim x .
(6.3.1)
x
k = lim xk, .
(6.3.2)
For any k {1, . . . , N }, we let

+
Once again, we pass to a subsequence so that these limits do exist. We clearly

have that S = {
xk }k=1,...,N where S is as in Theorem 4.1. We let also for any
k {1, . . . , N }
x
k, = expxk, (k, xk )
(6.3.3)
183
where xk Rn is given by Theorem 4.1. We obviously have that

lim x
k, = x
k
for all k {1, . . . , N }. We let k, be the bubble given by
n2 1
k,
k, (x) =
,
d (
xk, ,x)2
2
k, + gn(n2)
(6.3.4)
where x M and k {1, . . . , N }. It is easily checked that for any R > 0,

k,

0
(6.3.5)
vk, 1
L (Bxk, (Rk, ))
as +, where vk, is as in Theorem 4.1. Note also that there exists C > 1
independent of such that
1
k, k, Ck, in M
(6.3.6)
C
where k, is as in section 6.2, relation (6.2.1). In particular, the upper estimate of
i, s, and leave the i, s
Theorem 6.1 is also true if we replace the xi, s by the x
unchanged. Finally, we let S C 0 (M M ) be the function defined by
if x = y
1
S (x, y) =
(6.3.7)
n2
G (x, y)
if x = y
(n 2) n1 dg (x, y)
where (x, y) M M , and G is the Greens function of g + h . From the
property (P3) of Appendix A, S is indeed continuous on M M .
We assume in the following that the xi, s are ordered such that
1, 2, N, .
We define the sets i, , i = 1, . . . , N , in the following way:

1, = x M s.t. 1, (x) i, (x) , i = 1, . . . , N and

(6.3.8)
j, = x M \ j1
i=1 i, s.t. j, (x) i, (x) , i = 1, . . . , N
when j = 2, . . . , N . It is easily seen that M = N
i=1 i, and that the i, s are
disjoint. From (6.2.1) and (6.3.6), for any k {1, . . . , N } and any x k, ,
u (x) u0 (x) (1 + ) + Ck, (x)
(6.3.9)
where C > 0 is independent of , and ( ), independent of x, is a sequence of

positive real numbers converging to 0 as +.
6.3.1 The case u0 0. Since u0 0, we get with (6.3.9) that for any k such
that k {1, . . . , N }, and any x k, ,
u (x) Ck, (x) .
(6.3.10)
184
CHAPTER 6
We write with the Greens representation formula that
2 1
u (x ) =
G (x , x) u (x)
dvg
M

k=1
k,
G (x , x) u (x)
2 1
(6.3.11)
dvg
where G is the Greens function of g + h . We fix k {1, . . . , N }, and let

1
1
= x k, s.t. dg (x , x) dg (
xk, , x ) ,
Bk,
2

1
2
= x k, s.t. dg (x , x) < dg (
xk, , x ) .
Bk,
2
1
2
1
2
Clearly, k, = Bk,
Bk,
, and Bk,
Bk,
= . Given R > 0, we write that
2 1
G (x , x) u (x)
dvg = IR, + IIR, + IIIR,
(6.3.12)
k,
where
IR, =
IIR, =
IIIR, =
k, Bx
k, (Rk, )
1 \B
Bk,
x
k, (Rk, )
2 \B
Bk,
x
k, (Rk, )
G (x , x) u (x)
2 1
G (x , x) u (x)
2 1
G (x , x) u (x)
2 1
dvg ,
dvg ,
dvg .
Now we estimate these three integrals. We distinguish two cases.

Case 1. We assume that, up to a subsequence,
xk, , x )
dg (
+
k,
as +. We start by estimating IIIR, . From (6.3.10) and the definition
2
, noting that dg (
xk, , x) dg (
xk, , x ) dg (x , x), there exists C > 0
of Bk,
2
independent of such that, for any x Bk,
,
n

2 1 2
n

(
x
,
x)
d
1
1
g
k,
2 1
Ck, 2 1 +
u (x)
n (n 2)
2k,
n

2 1 2
(
x
,
x
)
d
1
1 n
g
k,
Ck, 2 1 +
4n (n 2)
2k,
n
+1
(n+2)
2
Ck,
dg (
xk, , x )
In particular,

n
2 +1
(n+2)
IIIR, = O k, dg (
xk, , x )
2
Bk,
G (x , x) dvg
185
From (6.1.80),

G (x , x) dvg = O
2
Bk,

2n
2
Bk,
dg (x , x)
dvg
=O

Bx
2n
dg x
k, ,x
2
xk, , x )
= O dg (

) dg (x , x)
dvg

.
Since dg (x , x
k, ) 1
k, + as +, we then get that

n
2n
2 1
.
dg (
xk, , x )
IIIR, = o k,
Coming back to the expression for k, (x ), it easily follows that, for any R > 0,

IIIR, = o k, (x ) .
(6.3.13)
1
Now we estimate IIR, . From (6.1.80) and the definition of Bk,
,
2n
G (x , x) C2 2n2 dg (
xk, , x )
1
for all x Bk,
. Together with (6.3.10) this gives that
2n
xk, , x )
IIR, Cdg (
1 \B
Bk,
x
k, (Rk, )
2 1
k,
dvg
where C > 0 is independent of and R. It is easily checked that
n
2 1
2 1
k,
dvg = R, k,
M \Bx
k, (Rk, )
where limR+ lim+ R, = 0. It follows that

IIR, = R, k, (x )
(6.3.14)
where R, is such that limR+ lim+ R, = 0. Finally, we estimate IR, .

xk, , x ) 1
Since dg (
k, + as +, we easily get (see Appendix A) that

k )
S (x , x
n2

dg (
0
xk, , x )
G (x , .)

(n 2) n1 L (Bx (Rk, ))
k,
as +. It follows that

k )
S (x , x
2n
IR, =
+ o (1) dg (
xk, , x )
(n 2) n1
2 1
u
dvg .
(6.3.15)
k, Bx
k, (Rk, )
Let i {1, . . . , N }, i = k, and assume that there exists a sequence of points (z )

in Bxk, (Rk, ) such that i, (z ) k, (z ). Then,

2
xi, , z )
dg (
1
R2
i, 1 +
1
+
k,
n (n 2)
2i,
n (n 2)
186
CHAPTER 6
so that
2
dg (
xi, , z ) n (n 2) 2i,
k,
i,

1+
R2
n (n 2)
$
1 .
xi, , z ) = O (k, ), and

This is possible if and only if i, = O (k, ). Then dg (
xk, , z ) = O (k, ), we have that dg (
xi, , x
k, ) = O (k, ). Using
since dg (
(4.1.6) and (6.3.3), we then obtain that i, = o (k, ), so that

2
xi, , z ) = O (i, k, ) = o 2k, .
dg (
This clearly proves that, for any R > 0,

Volg Bxk, (Rk, ) \k, = o nk, .
Writing that
2 1
u
dvg =
k, Bx
k, (Rk, )
(6.3.16)
2 1
u
dvg
Bx
k, (Rk, )
2 1
u
dvg
Bx
k, (Rk, )\k,
we get with (6.3.16), and from Holders inequalities that
2 1
2 1
u
dvg =
u
dvg
k, Bx
k, (Rk, )
Bx
k, (Rk, )
+O

1
Volg Bxk, (Rk, ) \k, 2
n
2 1
2 1
.
u
dvg + o k,
Bx
k, (Rk, )
By (6.3.3), for R large,

1
Bxk, ( Rk, ) Bxk, (Rk, ) Bxk, (2Rk, ) .
2
From point (P1) of Theorem 4.1, we then get that

1 n
2 1
lim k,2
u
dvg =
u2 1 dx
lim
R+ +
Bx
k, (Rk, )
Rn
where u is as in (4.1.5). Since
n

n1
(n (n 2)) 2
u2 1 dx =
n
n
R
this leads to
Bx
k, (Rk, )
2 1
u
dvg =
n1
n
n
2 1
(n (n 2)) 2 + R, k,
where limR+ lim+ R, = 0. Coming back to (6.3.15), we then get that

n

n
1
2n
2 1
xk , x ) + R, k,
dg (
xk, , x )
.
IR, = (n (n 2)) 2 S (
187
xk, , x ) 1
Since dg (
k, + as +, we also have that
lim
k, (x )
n
2 1
2n
k, dg (
xk, , x )
= (n (n 2)) 2
Hence,
xk , x ) + R, ) k, (x ) .
IR, = (S (
(6.3.17)
Combining (6.3.12) with (6.3.13), (6.3.14), and (6.3.17), it follows that

2 1
G (x , x) u (x)
dvg = S (
xk , x ) + o (1) k, (x ) . (6.3.18)
k,
This ends case 1.

dg (
xk, , x )
Rk
k,
as + for some Rk 0. We start by estimating IIIR, . By the definition of
2
2
, for any x Bk,
,
Bk,
xk, , x) dg (
xk, , x ) + dg (x , x)
dg (
3
xk, , x ) .
dg (
2
Hence, for large,
2
Bxk, (2Rk k, ) .
Bk,
It follows that
2
\Bxk, (Rk, ) =
Bk,
when is large and R is sufficiently large. In particular,

lim
lim IIIR, = 0 .
R+ +
1
Now we estimate IIR, . For any x Bk,
\Bxk, (Rk, ), we have that
xk, , x) dg (x , x
k, )
dg (x , x) dg (
xk, , x) (Rk + o (1)) k,
dg (
(R Rk + o (1)) k, .
1
\Bxk, (Rk, ),
Thus, for large, for R large, and for any x Bk,
1
dg (
xk, , x) .
2
Then we can write, using (6.1.80) and (6.3.10), that
2n
2 1
dg (
xk, , x)
k, (x)
dvg
IIR, C
dg (x , x)
1 \B
Bk,
x
k, (Rk, )
(6.3.19)
188
CHAPTER 6
where C > 0 is independent of and R. Direct computations give that
n
2n
2 1
2 1
lim k,
dg (
xk, , x)
k, (x)
dvg = 0
lim
R+ +
M \Bx
k, (Rk, )
so that
1 n
IIR, = R, k,2
where limR+ lim+ R, = 0. Since dg (

xk, , x ) 1
k, Rk as the
parameter +, we also have that

1 n2
2
R
1 n
k
+ o (1) .
1+
k, (x ) = k,2
n (n 2)
In particular, it follows that
IIR, = R, k, (x )
(6.3.20)
where limR+ lim+ R, = 0. Finally, we estimate IR, . We let

1
exp1
(6.3.21)
xk, (x ) .
k,

Given 0 < < ig /2, and for x B0 1
k, , the Euclidean ball of center 0 and
z =
radius 1
k, , we also let

n
2 1
u expxk, (k, x) ,
uk, (x) = k,
gk, (x) = expxk, g (k, x) .

Then we can write that
n
2
IR, = k,
where
Vk, (R)
2
G (x)uk, (x)
dvgk,

G (x) = G expxk, (k, z ) , expxk, (k, x)
and
Vk, (R) =
1
k,

exp1
k, (Rk, ) .
xk, k, Bx
Using the property (P3) of Appendix A, for any x Vk, (R),

G expxk, (k, z ) , expxk, (k, x)

2n
1 + o (1)
=
dg expxk, (k, z ) , expxk, (k, x)
(n 2) n1
1 + o (1)
2n
=
2n dg (z , x)
.
(n 2) n1 k, k,
Hence,
1 + o (1)
1 n
2n
2 1
k,2
1Vk, (R) dgk, (z , x)
uk, (x)
dvgk,
IR, =
(n 2) n1
Rn
189
where 1Vk, (R) is the characteristic function of Vk, (R). It is easily seen that
2
(Rn ) as +, where is the Euclidean metric. Using (6.3.3)
gk, in Cloc
and (6.3.16), we have that
1Vk, (R) 1Bxk (R) a.e.
as +. From point (P1) of Theorem 4.1, we have that
uk, u (. xk ) a.e.
as +. Finally, up to a subsequence,
dgk, (z , .)
2n
|z .|2n a.e.

as +, where z = lim+ z . By (6.3.10), 1Vk, (R) uk, is uniformly bounded in Rn . Using the Lebesgue-dominated convergence theorem we
then easily get that

1
1 n
2 1
IR, = k,2
|z x|2n u (x xk )
dx + R, .
(n 2) n1 Rn
Since u = u2
, we have that
2 1
|z x|2n u (x xk )
dx = u (z xk ) .
1
(n 2) n1
Rn

n
2
k,
k, (x ) u (z xk )
as +. Hence, we have proved that

IR, = (1 + R, ) k, (x ) .
(6.3.22)
Combining (6.3.12) with (6.3.19), (6.3.20), and (6.3.22), we then get that
2 1
G (x , x) u (x)
dvg = (1 + o (1)) k, (x ) .
(6.3.23)
k,
This ends case 2.

From cases 1 and 2, and since S (x, x) = 1,

2 1
G (x , x) u (x)
dvg = S (
xk , x ) + o (1) k, (x )
k,
for all k {1, . . . , N }. By (6.3.11), we then get that

u (x ) =
N

xk , x ) + o (1) k, (x ) .
(S (
(6.3.24)
k=1
We already know from the preceding section and (6.3.6) that there exists C > 1
such that, for any x M ,
u (x) C
N

k=1
k, (x) .
190
CHAPTER 6
Let C > 1 be sufficiently large such that S (x, y) 2/C for all x, y M . Since
S is positive and continuous, such a C exists. Then, using (6.3.24), we get that,
for large,
u (x)
N
1
k,
C
k=1
and this proves Theorem 6.1. As a remark, the xi, s of Theorem 6.1 are not those
of Theorem 4.1, but the x
i, s defined in (6.3.3). Summarizing, Theorem 6.1 is
proved in the special case u0 0.
From standard elliptic theory, it easily follows from (6.3.24) that, up to a subsequence,
n 1
N

i, 2
1 n
lim 1, 2 u = Cn
G (
xi , .)
(6.3.25)
lim
+
+ 1,
i=1
2
(M \S), where
in Cloc
n
Cn = (n (n 2)) 2
(n 2) n1 ,
and S is as in Theorem 4.1.

6.3.2 The case u0 0. Following (6.3.9), we have that for any k {1, . . . , N }
and any x k, ,

2 1
2 1
u (x)
C k, (x)
+1
(6.3.26)
where C > 0 is independent of . We write with the Greens representation formula that

2 1
2 1
0
u (x ) u (x ) =
dvg
G (x , x) u (x)
u0 (x)
M
+
G (x , x) (h (x) h (x)) u0 (x) dvg .
M
By (6.0.1) and (6.1.80), we get that
G (x , x) (h (x) h (x)) u0 (x) dvg = o (1)

M
so that
u (x ) u0 (x ) =

2 1
2 1
dvg + o (1) .
G (x , x) u (x)
u0 (x)
By point (P1) of Theorem 4.1, after passing to a subsequence, there exists 0

as + such that
sup
M \N
i, ( )
i=1 Bx
|u u0 | 0
as +. Then it follows from (6.1.80) that

2 1
2 1
G (x , x) u (x)
u0 (x)
dvg = o (1)
M \N
i, ( )
i=1 Bx
191
and that
N
i, ( )
i=1 Bx
G (x , x) u0 (x)
2 1
dvg = o (1) .
Hence,
u (x ) u0 (x )
N

(6.3.27)
2 1
=
dvg + o (1) .
G (x , x) u (x)
N
i, ( ))
k=1 k, (i=1 Bx
We fix k {1, . . . , N }, and let

N

1
1
Bk, = x k, i=1 Bxi, ( ) s.t. dg (x , x) dg (
xk, , x ) ,
2

N

1
2
xk, , x ) .
Bk, = x k, i=1 Bxi, ( ) s.t. dg (x , x) < dg (
2
1
2
1
2
Bk,
and Bk,
Bk,
= . Up to replacing by
Clearly, k, = Bk,

= max , ( max i, )1/2 ,
i=1,...,N
we can assume that for any R > 0, and any k {1, . . . , N },

Bxk, (Rk, ) Bxk, ( ) .
Given R > 0, we can then write that for large,
2 1
dvg
G (x , x) u (x)
k, (N
i, ( ))
i=1 Bx
(6.3.28)
= IR, + IIR, + IIIR, ,

where
IR, =
k, Bx
k, (Rk, )
G (x , x) u (x)
IIR, =
1 \B
Bk,
x
k, (Rk, )
G (x , x) u (x)
2 1
G (x , x) u (x)
2 1
IIIR, =
2 \B
Bk,
x
k, (Rk, )
2 1
dvg ,
dvg ,
dvg .
Now we estimate these three integrals. Here again we distinguish two cases.
k, )
dg (x , x
+
k,
as +. First we estimate IIIR, . From (6.3.26) and the definition of
2
, noting that dg (
xk, , x) dg (
xk, , x ) dg (x , x), there exists C > 0
Bk,
2
independent of such that for any x Bk,
,
n

2 1 2
n

(
x
,
x
)
d
1
1
g
k,
2 1
C k, 2 1 +
+ 1
u (x)
4n (n 2)
2k,

n
(n+2)
2 +1
dg (
xk, , x )
+1 .
C k,
192
CHAPTER 6
It follows that

n
2 +1
(n+2)
xk, , x )
IIIR, = O k, dg (

+O
G (x , x) dvg
By (6.1.80),
2
Bk,
2
Bk,
2
Bk,

2n
2
Bk,
G (x , x) dvg
G (x , x) dvg = O
dg (x , x)
dvg
=O

Bx
2n
dg x
k, ,x
xk, , x )
= O dg (

) dg (x , x)
dvg

.
2
N
Since 0 and Bk,
i=1 Bxi, ( ), it also follows from (6.1.80) that
G (x , x) dvg = o (1) .
2
Bk,
Since dg (x , x
k, ) 1
k, + as +, this gives that
n

2n
2 1
dg (
xk, , x )
+ o (1) .
IIIR, = o k,
Coming back to the definition of k, , it easily follows that

IIIR, = o k, (x ) + o (1)
(6.3.29)
1
,
for all R > 0. Now we estimate IIR, . From (6.1.80) and the definition of Bk,
2n
G (x , x) C2 2n2 dg (
xk, , x )
1
for all x Bk,
. We can then write with (6.3.26) that
2n
IIR, Cdg (
xk, , x )
1 \B
Bk,
x
k, (Rk, )
+C
1 \B
Bk,
x
k, (Rk, )
2 1
k,
dvg
G (x , x) dvg
where C > 0 is independent of and R. It is easily checked that
n
2 1
2 1
k,
dvg = R, k,
M \Bx
k, (Rk, )
where limR+ lim+ R, = 0. As above, since 0 as + and

1
N
Bk,
i, ( ), we also get with (6.1.80) that
i=1 Bx
G (x , x) dvg = o (1) .
1
Bk,
Hence, coming back to the definition of k, ,

IIR, = R, k, (x ) + 1 .
193
(6.3.30)
Finally, we estimate IR, . It is easily checked that the arguments used to prove
(6.3.17) can be used here. Following the proof of (6.3.17) we then get that

xk , x ) + R, k, (x ) .
(6.3.31)
IR, = S (
Combining (6.3.28) and (6.3.29)(6.3.31), it follows that
2 1
dvg
G (x , x) u (x)
k, (N
i, ( ))
i=1 Bx

= S (
xk , x ) + o (1) k, (x ) + o (1) .
(6.3.32)
This ends case 1.

xk, , x )
dg (
Rk
k,
as + for some Rk 0. First we estimate IIIR, . By the definition of
2
2
, for any x Bk,
,
Bk,
xk, , x) dg (
xk, , x ) + dg (x , x)
dg (
3
xk, , x ) .
dg (
2
Hence, for large,
2
Bxk, (2Rk k, ) .
Bk,
It follows that
2
\Bxk, (Rk, ) =
Bk,
when is large and R is sufficiently large. In particular,

lim
lim IIIR, = 0 .
R+ +
1
Now we estimate IIR, . For any x Bk,
\Bxk, (Rk, ), we have that
xk, , x) dg (x , x
k, )
dg (x , x) dg (
xk, , x) (Rk + o (1)) k,
dg (
(R Rk + o (1)) k, .
1
\Bxk, (Rk, ),
Thus, for large, for R large, and for any x Bk,
1
dg (
xk, , x) .
2
Then we can write, using (6.1.80) and (6.3.26), that
2n
2 1
dg (
xk, , x)
k, (x)
dvg
IIR, C
dg (x , x)
1 \B
Bk,
x
k, (Rk, )
+C
1
Bk,
G (x , x) dvg
(6.3.33)
194
CHAPTER 6
where C > 0 is independent of and R. Direct computations give that
n
2n
2 1
2 1
lim k,
dg (
xk, , x)
k, (x)
dvg = 0 .
lim
R+ +
M \Bx
k, (Rk, )
Moreover, since dg (
xk, , x ) 1
k, Rk as +, we also have that

1 n2
2
R
1 n
k
+ o (1) .
k, (x ) = k,2
1+
n (n 2)
1
Independently, since 0 and Bk,
N
i, ( ), it also follows from
i=1 Bx
(6.1.80) that
G (x , x) dvg = o (1) .
1
Bk,
In particular, it follows that

IIR, = R, k, (x ) + 1 .
(6.3.34)
Finally, we estimate IR, . It is easily checked that the arguments used to prove
(6.3.22) can be used here. Following the proof of (6.3.22) we then get that

IR, = 1 + R, k, (x ) .
(6.3.35)
Combining (6.3.28) with (6.3.33)(6.3.35), it follows that
2 1
dvg
G (x , x) u (x)
k, (N
i, ( ))
i=1 Bx
(6.3.36)
= (1 + o (1)) k, (x ) + o (1) .
This ends case 2.
From cases 1 and 2, and since S (x, x) = 1,
2 1
dvg
G (x , x) u (x)
k, (N
i, ( ))
i=1 Bx
= (S (
xk , x ) + o (1)) k, (x ) + o (1)
for all k {1, . . . , N }. By (6.3.27), we then get that
u (x ) = u0 (x ) (1 + o (1))
+
N

xk , x ) + o (1) k, (x ) .
(S (
(6.3.37)
k=1
We already know form the preceding section and (6.3.6) that there exists C > 1
such that, for any x M ,
u (x) u0 (x) (1 + o (1)) + C
N

k=1
k, (x) .
195
Let C > 1 be sufficiently large such that S (x, y) 2/C for all x, y M . Since
S is positive and continuous, such a C exists. Then, thanks to (6.3.37), we get
that, for large,
N
1
k, (x)
u (x) u (x) (1 + o (1)) +
C
0
k=1
and this proves Theorem 6.1. As a remark, the xi, s of Theorem 6.1 are not those
of Theorem 4.1, but the x
i, s defined in (6.3.3). Summarizing, Theorem 6.1 is
proved in the special case u0 0. Together with subsection 6.3.1, Theorem 6.1 is
proved.
The remark following Theorem 6.1 easily follows from (6.3.37) and the fact that
S (x, x) = 1 for all x. We prove now that the bubbles of Theorem 6.1 satisfy
Theorem 3.1. More precisely, we prove that, up to a subsequence,
u = u0 +
N
i, + o (1)
(6.3.38)
i=1
where o (1) H12 (M ) 0 as +, and that

N
Ig (u ) = Ig u0 + Knn + o (1)
n
where
Ig (u) =
1
2
and
Ig (u) =
1
2

1
|u|2 + h u2 dvg
2

1
|u|2 + h u2 dvg
2
(6.3.39)
|u|2 dvg
|u|2 dvg .
Since
2 1
g u + h u = u
in M , we have that
Ig
1
(u ) =
n
u2 dvg .
Since
g u0 + h u0 = (u0 )2
in M , we also have that

1
Ig u0 =
n
(u0 )2 dvg .
It follows that the proof of (6.3.39) reduces to the proof that

u2 dvg =
(u0 )2 dvg + N Knn .
lim
+
(6.3.40)
196
CHAPTER 6
Using the asymptotic estimates (6.3.24) and (6.3.37) of this section,

lim
u2 dvg
+
= lim
u (x) +
2
N
i, (x) S (
xi , x)
(6.3.41)
dvg .
i=1
Since S is continuous on M M ,

u0 (x) +
N
2
i, (x) S (
xi , x)
i=1
(u0 )2 dvg +
N
(u0 )2
+O
M
+O
i, (x) S (
xi , x)

dvg
2 1
i,
dvg
(6.3.42)
i,
i=1
N
2
i=1
N
dvg
dvg
i=1
Direct computations give that
i, dvg 0 and
M
2 1
i,
dvg 0
as +, for all i {1, . . . , N }. By (6.3.42) we then get that

u0 (x) +
N
2
i, (x) S (
xi , x)
i=1
(u0 )2 dvg +
N
dvg
(6.3.43)
2
i, (x) S (
xi , x)
dvg + o (1) .
i=1
We write now that

N
2
i, (x) S (
xi , x)
dvg
i=1

i=1
i, (x)
2
2
S (
xi , x)
+O
i=j
dvg
2 1
i, j,
dvg .
(6.3.44)
197
Let i {1, . . . , N }. Given R > 0, we write that
2
2
i, (x) S (
xi , x) dvg
M
2
2
=
i, (x) S (
xi , x) dvg
Bx
i, (Ri, )
M \Bx
i, (Ri, )
i, (x)
2
2
S (
xi , x)
= (1 + o (1))
Bx
i, (Ri, )
M \Bx
i, (Ri, )
i, (x)
i, (x)
2
2
dvg
dvg
2
S (
xi , x)
dvg .
Here we used that S is continuous on M M and that S (x, x) = 1 for all

x M . Direct computations give that
2
lim
i, (x) dvg = Knn
lim
R+ +
and that
Bx
i, (Ri, )
lim
lim
R+ +
Thus,
M \Bx
i, (Ri, )
lim
i, (x)
2
i, (x)
2
S (
xi , x)
2
dvg = 0 .
dvg = Knn
(6.3.45)
for all i {1, . . . , N }. Let i, j {1, . . . , N }, i = j, and R > 0. Assume for

instance that i, j, . We write that

2 1
2 1
i, j,
dvg
+ j, i,
M

2 1
2 1
i, j,
dvg
+ j, i,
Bx
j, (Rj, )
+

M \Bx
j, (Rj, )
2
j,
dvg
Bx
j, (Rj, )
Bx
j, (Rj, )
M \Bx
j, (Rj, )
2
i,
dvg
+
M
221
2
i,
dvg

2 1
2 1
+ j, i,
i, j,
dvg
Bx
j, (Rj, )
221

Bx
j, (Rj, )
2
j,
dvg
2 1
2
2
i,
dvg
221

21
2
j,
dvg
2
i,
dvg
21
M \Bx
j, (Rj, )
2
j,
dvg
21
.
21
198
CHAPTER 6
As above,
M \Bx
j, (Rj, )
where lim
2
j,
dvg = R,
lim R, = 0. Since (i, 2 ) and (j, 2 ) are bounded, we
R+ +
get that
2 1
2 1
i, j,
+ j, i,
dvg = O
Bx
j, (Rj, )
21
2
+ R,
i,
dvg

+O
If
dg (xi, ,xj, )
i,
221

2
.
i,
dvg
Bx
j, (Rj, )
+ as +, we have that
Bx
j, (Rj, )

2n n
2
i,
dvg = O ni, dg (xi, , xj, )
j,
= o (1) .
If dg (xi, , xj, ) = O (i, ), we then have by (4.1.6) that j, = o (i, ) and thus
that

2
n
i,
dvg = O n
i, j,
Bx
j, (Rj, )
= o (1) .
Therefore,
2 1
i, j,
dvg 0
(6.3.46)
as +, for all i, j {1, . . . , N }, i = j. Combining (6.3.41) and (6.3.43)

(6.3.46), we then get that (6.3.40) holds. As already mentioned, this proves that
(6.3.39) holds. We now let
R = u u 0
N
i, .
i=1
The proof of (6.3.38) reduces to the proof that, up to a subsequence,
2
|R |2 dvg 0 and
R
dvg 0
M
(6.3.47)
as +. Up to a subsequence, we may assume that u u0 in L2 (M ) as

2
dvg 0 as + for all i, we then get that
+. Noting that M i,
2
dvg 0
(6.3.48)
R
M
199
as +. Independently, we have that
|R |2 dvg
M
|u |2 dvg +
i=j

i=1
|i, |2 dvg

u u0 dvg
(i, j, ) dvg 2

i=1
|u0 |2 dvg +
(6.3.49)
(u i, ) dvg + 2

u0 i, dvg .
i=1
It easily follows from (6.3.40) that
2
|u | dvg
|u0 |2 dvg + N Knn
(6.3.50)
as +, while it is clear that

|u0 |2 dvg ,
u u0 dvg
M
M
0

u i, dvg 0 ,
M
|i, |2 dvg Knn
(6.3.51)
as +, for all i. From (6.3.46) we can also prove that
(i, j, ) dvg 0
(6.3.52)
for all i = j. Independently,
(u i, ) dvg =
(g u )i, dvg
M
so that, from the equations satisfied by u ,
2 1
i, dvg + o(1) .
(u i, ) dvg =
u
M
Given R > 0, we write that
2 1
u
i, dvg
M
2 1
=
u
i, dvg +
Bx
i, (Ri, )
Since
M \Bx
i, (Ri, )
lim
lim
R+ +
M \Bx
i, (Ri, )
2 1
i, dvg .
u
i, (x)
2
dvg = 0 ,
200
CHAPTER 6
it follows that
2 1
u
i, dvg
where
lim
Bx
i, (Ri, )
2 1
u
i, dvg + R,
lim R, = 0. Independently, we easily get with (6.3.46) and
R+ +
using the asymptotics (6.3.24) and (6.3.37), that
2 1
u
i, dvg = Knn + R,
Bx
i, (Ri, )
where R, is as above. Hence,
(u i, ) dvg Knn
(6.3.53)
as +. Combining (6.3.49)(6.3.53), it follows that
|R |2 dvg 0
(6.3.54)
as +. Using (6.3.48) and (6.3.54), we get that (6.3.47) holds. As already

mentioned, this proves that (6.3.38) also holds.
Appendix A
The Greens Function on Compact Manifolds
We discuss in this appendix existence and properties of Greens functions for operators like g + h. We let (M, g) be a smooth compact Riemannian manifold of
dimension n 3, and let h C 0, (M ), 0 < < 1. We assume that g + h is
coercive, and let > 0 be such that for any u H12 (M ),

2
2
(A1)
|u| + hu dvg
|u|2 + u2 dvg .
M
Given K > 0 and > 0, we let H(K, ) be the set of the functions h in C 0, (M )
which are such that (A1) is satisfied and
|h(x)| K for all x ,
|h(y) h(x)| Kdg (x, y) for all x, y .
(A2)
Let DM be the subset of M M consisting of the (x, x), x M . Let h H(K, )

for some K, > 0. We claim in this appendix that there exists a continuous function G : M M \DM R such that for any x M , Gx = G(x, .) is in L1 (M ),
such that for any u C 2 (M ) and any x M ,
G(x, y) (g u(y) + h(y)u(y)) dvg (y) ,

(A3)
u(x) =
M
and such that the following holds:

(P1) For any x, y M , x = y, G(x, y) = G(y, x), and G(x, y) > 0. Moreover,
for any x M , the function Gx : M \{x} R given by Gx (y) = G(x, y), is in
2,
(M \{x}).
Cloc
(P2) For any x, y M , x = y, dg (x, y)n2 G(x, y) C, where C > 0 depends
only on (M, g), K and .
(P3) There exists > 0, depending only on (M, g), such that for any x, y M ,
x = y, if dg (x, y) < , then

1
Cdg (x, y) , and
dg (x, y)n2 G(x, y)

(n 2)n1

dg (x, y)n1 |y G(x, y)| 1 Cdg (x, y)

n1
where n1 is the volume of the unit (n 1)-sphere, and C > 0 depends only on
(M, g), K and .
We prove the existence of G and (P1)(P3) in what follows. We let > 0
be such that < ig /3, where ig is the injectivity radius of (M, g). We also let
202
APPENDIX A
C (M M ), 0 1, be such that (x, y) = 1 if dg (x, y) , and

(x, y) = 0 if dg (x, y) 2. For x = y, we set
H(x, y) =
x (y)
(n 2)n1 dg (x, y)n2
where x (y) = (x, y), and n1 is the volume of the unit (n 1)-sphere. Following Aubin [7], it is easily seen that there exists C > 0 such that for any x, y,
x = y,
C
,
dg (x, y)n2
C
|g,y H(x, y)|
dg (x, y)n2
|H(x, y)|
(A4)
and such that

g,y,dist. H(x, y) = x + g,y H(x, y)
in the sense of distributions, where x is the Dirac measure at x. Letting
f (x, y) = g,y H(x, y) ,
we thus have that for any C 2 (M ), and any x M ,
H(x, y)g (y)dvg (y)

f (x, y)(y)dvg (y) .
(x) =
M
(A5)
Let X, Y : M M \DM R be continuous functions such that for any x = y,

|X(x, y)| CX dg (x, y)n and |Y (x, y)| CY dg (x, y)n
where CX , CY > 0, and 0 < , < n are independent of x and y. For x = y, we
let
Z(x, y) =
X(x, z)Y (z, y)dvg (z) .
M
From a result by Giraud [38] (see also Aubin [7]) Z : M M \DM R is continuous and there exists C(, , M, g) > 0, depending only on , , and (M, g),
such that
|Z(x, y)| CX CY C(, , M, g)dg (x, y)+n if + < n ,
|Z(x, y)| CX CY C(, , M, g) (1 + | ln dg (x, y)|) if + = n , and (A6)
|Z(x, y)| CX CY C(, , M, g) if + > n .
Moreover, when + > n, Z is continuous on M M . We let 1 be the function
from M M \DM to R given by
1 (x, y) = g,y H(x, y) h(y)H(x, y)
(A7)
and then, by induction, we define i : M M \DM R by
i (x, z)1 (z, y)dvg (z) .

i+1 (x, y) =
(A8)
THE GREENS FUNCTION ON COMPACT MANIFOLDS
We also let
203
n
+1
2
where E 2 is the greatest integer not exceeding n2 . It is easily seen from (A4)
and (A6) that for h H(K, ), any i 1, . . . , k + 1, and any x = y,
k=E
n
|i (x, y)|
C
if 2i < n ,
dg (x, y)n2i
|i (x, y)| C (1 + | ln dg (x, y)|) if 2i = n , and

|i (x, y)| C if 2i > n ,
(A9)
where C > 0 depends only on (M, g) and K. We also have that k and k+1 are
continuous on M M . Given x M , we let ux be given by
g ux + hux = k+1,x
(A10)
where k+1,x (y) = k+1 (x, y). Since g + h is coercive, ux exists. We then let
G : M M \DM R be given by
k

G(x, y) = H(x, y) +
i (x, z)H(z, y)dvg (z) + u(x, y)
(A11)
i=1
where u(x, y) = ux (y). It is easily checked that for any x M , Gx = G(x, .) is

C 2 (M ) and any x M ,
in L1 (M ), and that for any u
u
(x) =
G(x, y) (g u
(y) + h(y)
u(y)) dvg (y) .
M
In particular, (A3) is satisfied. Now we claim that there exists C > 0, depending
only on (M, g) and K, such that for any x, y, y M ,
|k+1 (x, y ) k+1 (x, y)| Cdg (y, y ) .
(A12)
In order to prove this claim, we proceed as follows. It is easily checked from (A7)
and (A9) that
|k+1 (x, y ) k+1 (x, y)|

Cdg (y, y ) + C
|f (z, y ) f (z, y)| dvg (z)
M

1
1

+C
dg (z, y)n2 dg (z, y )n2 dvg (z)
M
(A13)
where C > 0 depends only on (M, g) and K. We let > 0 to be chosen later on,
and assume that dg (y, y ) < /2. We can write that

1
1

dvg (z)
dg (z, y)n2

n2
dg (z, y )
M

1
1

dvg (z)
=
(A14)

n2

n2
dg (z, y )
By ( ) dg (z, y)

1
1

dvg (z) .
+

n2
dg (z, y )n2
M \By ( ) dg (z, y)
204
APPENDIX A
If dg (z, y) , then dg (z, y ) /2. It easily follows that

1
1

dvg (z) Cdg (y, y )

n2
dg (z, y )n2
M \By ( ) dg (z, y)
(A15)
where C( ) > 0 depends only on (M, g), , and . Similarly, we can write that

n3

1
1
(n 2) (dg (z, y) + dg (z, y ))

dg (y, y )
dg (z, y)n2
dg (z, y )n2
dg (z, y)n2 dg (z, y )n2
so that

1
1

dvg (z)

n2

n2
dg (z, y )
By ( ) dg (z, y)
(A16)
(n 2)I(y, y )dg (y, y )

where
I(y, y ) =
(dg (z, y) + dg (z, y ))

dvg (z) .
n3
By ( )
We choose > 0 sufficiently small such that for any x M , and any s, t Rn , if
|s| and |t| , then
1
|t s| dg (expx (s), expx (t)) C( )|t s|
C( )
where C( ) > 1 does not depend on x, s, and t. We let z = expy (t) and let
y = expy (t0 ), |t0 | < /2. Then easy computations give that
n3
(|t| + |t t0 |)
dt C( )
(A17)
I(y, y ) C( )
n2 |t t |n2
|t|

0
B0 ( )
where C( ) > 0 depends only on (M, g) and . Combining (A14)(A17), we
proved that there exists C > 0, depending only on (M, g), such that for any points
y, y M , if dg (y, y ) < /2, then

1
1

dvg (z) Cdg (y, y ) .
(A18)
dg (z, y)n2

n2
d
(z,
y
)
g
M
Clearly, (A18) extends to all y, y M , so that, from (A13) and (A18),
|k+1 (x, y ) k+1 (x, y)|

Cdg (y, y ) + C
|f (z, y ) f (z, y)| dvg (z)
(A19)
for all y, y M , where C > 0 depends only on (M, g) and K. We can write (see,
for instance, Aubin [7]) that
f (x, y) =
Q(x, y)
dg (x, y)n1
where, in geodesic normal coordinates at x,

1
Qx =
r ln |g|
n1
205
and Qx (y) = Q(x, y). In particular, there exists C > 0, depending only on (M, g),
such that for any x, y, y ,
|Q(x, y)| Cdg (x, y) and |Q(x, y ) Q(x, y)| Cdg (y, y ) .
It follows that
|f (z, y ) f (z, y)| dvg (z) Cdg (y, y )

1
1

dvg (z) .
+C
dg (z, y)
n1
)n1
d
(z,
y)
d
(z,
y
g
g
M
Writing that

n2

1
1

(n 1) (dg (z, y) + dg (z, y ))
dg (y, y )
dg (z, y)n1
dg (z, y )n1
we get, as when proving (A18), that for any y, y M ,

1
1

dvg (z) Cdg (y, y )
dg (z, y)
dg (z, y)n1
dg (z, y )n1
(A20)
where C > 0 depends only on (M, g). Combining (A19) and (A20), it follows that
there exists C > 0, depending only on (M, g) and K, such that for any x, y, y in
M,
|k+1 (x, y ) k+1 (x, y)| Cdg (y, y ) .
This proves (A12). Now that we have (A12), we return to the definition (A10) of
ux . Clearly, ux C 2, (M ). Thanks to standard elliptic estimates as in GilbargTrudinger [37], see for instance Theorems 6.2 and 8.17, we have that

ux C 2, C ux C 0 + k+1,x C 0,
and

ux C 0 C ux 2 + k+1,x C 0
where C > 0 depends only on (M, g) and K. Since h H(K, ),

C
k+1,x C 0
and we get, using (A9), (A12), and the above equations, that for any x M ,
ux 2
ux C 2, C
(A21)
where C > 0 depends only on (M, g), K, and . Coming back to the definition
(A11) of G, and combining (A4), (A6), (A9), and (A21), we get that there exists
C > 0, depending only on (M, g), K, and , such that for any x, y M ,
dg (x, y)n2 G(x, y) C .
This proves (P2). Noting that
g (ux ux ) + h(ux ux ) = k+1,x k+1,x
we get, thanks to standard elliptic estimates as above, that
ux ux C 0 Ck+1,x k+1,x C 0
(A22)
206
APPENDIX A
where C > 0 depends only on (M, g), K, and . Writing that

|u(x , y ) u(x, y)| |ux (y) ux (y)| + |ux (y) ux (y )|
ux ux C 0 + ux C 0 dg (y, y )
we then get with (A21)(A22) that, for any x, x , y, y M ,
|u(x , y ) u(x, y)| C (k+1,x k+1,x C 0 + dg (y, y ))
where C > 0 depends only on (M, g), K, and . In particular, u is continuous on
M M . As an easy consequence, coming back to the definition (A11) of G, it
follows that G is continuous on M M \DM . Noting that
g Gx + hGx = 0
(A23)
2,
(M \{x}). Given x, y M such that
in M \{x}, we also have that Gx is in Cloc
0 < dg (x, y) < 1, we get with (A4), (A6), (A9), and (A21), that
|G(x, y) H(x, y)| Cdg (x, y)3n
(A24)
where C > 0 depends only on (M, g), K, and . In particular, if dg (x, y) ,

then

1
2n
Cdg (x, y) .
Gx (y) dg (x, y)
(n 2)n1
It follows that there exists > 0 such that for any x, y M , if 0 < dg (x, y) < ,
then Gx (y) 0. Let Gx be the nonpositive part of Gx . According to what we
just said, G
x has its support in M \Bx (). Moreover, Gx is Lipschitz so that
2
Gx H1 (M ). Using (A23) we can then write
0=
G
x (g Gx + hGx ) dvg
M

2
2
=
|G
dvg .
x | + h(Gx )
M
Since g + h is coercive, it follows that G

x 0, and we have that Gx 0 in
M \{x}. The Hopf maximum principle then gives that Gx is positive in M \{x}.
Independently, by standard arguments, we get that for any C (M ),
(G(x, y) G(y, x)) (y)dvg (y) = 0 .

M
It follows that G(x, y) = G(y, x) for all x = y, and (P1) is proved. We are thus
left with the proof of (P3). The first equation in (P3) is an easy consequence of
(A24). It remains to prove the second equation in (P3). We fix x, y M such that
0 < dg (x, y) < , and let y Rn be such that y = expx (
y ). Given X S n1 ,
n
the unit sphere of R , and t R, we set
y + tX) .
yt = expx (
Then,

d
(
y , X)
H(x, yt )t=0 =
dt
n1 dg (x, y)n1 |
y|
(A25)
where (., .) is the Euclidean scalar product, and we also have that

d
C
H(z, yt )
dt
t=0
dg (z, y)n1
207
(A26)
where C > 0 depends only on (M, g). For any r > 0, and any i = 1, . . . , k, from
(A26),

d

i (x, z)H(z, yt )dvg (z)
dt
t=0
M \By (r)
(A27)

d

H(z, yt )
dvg (z) .
=
i (x, z)
dt
t=0
M \By (r)
Independently, for r > 0 small, r < dg (x, y)/2, and t > 0 small, we can write that

H(z, yt ) H(z, y)

dvg (z)
i (x, z)

By (r)

t

dg (z, yt )2n dg (z, y)2n

C
dvg (z)

dg (x, y)n2 By (r)
t
n3
(dg (z, yt ) + dg (z, y))
C
dvg (z)
dg (x, y)n2 By (r) dg (z, yt )n2 dg (z, y)n2

where C > 0 depends only on (M, g). As when proving (A18), it follows that

H(z, yt ) H(z, y)

dvg (z) C(r)
i (x, z)
(A28)

By (r)

t
where C(r) 0 as r 0. Combining (A27) and (A28) we then get that for any
i = 1, . . . , k,

d

i (x, z)H(z, yt )dvg (z)
dt
t=0
M

(A29)

d

H(z, yt )
=
i (x, z)
dvg (z) .
dt
t=0
M
Using (A25) and (A29), coming back to the definition (A11) of G, we can write
that

d
G(x, yt )t=0
dt
(
y , X)
=
(A30)
n1 dg (x, y)n1 |
y|

k

d
d
+
H(z, yt )t=0 dvg (z) + ux (yt )t=0 .
i (x, z)
dt
dt
i=1 M
By (A21),

d
ux (yt ) C
dt
t=0
(A31)
208
APPENDIX A
where C > 0 depends only on (M, g), K, and . Independently, as an easy consequence of (A6), (A9), and (A26), for any i = 1, . . . , k,

C
d

H(z, yt ) t=0 dvg (z)
(A32)
i (x, z)

dt
d
(x,
y)n2
g
M
where C > 0 depends only on (M, g) and K. In geodesic normal coordinates at x,
we refer for instance to Lemma 8 in Hebey-Vaugon [47], |g ij (y) ij | Cdg (x, y)
where C > 0 can be chosen such that it depends only on (M, g). Combining
(A30)(A32), we then get that

n1 dg (x, y)n1 |y G(x, y)| 1 Cdg (x, y)
where C > 0 depends only on (M, g), K, and . This ends the proof of (P3).
Appendix B
Coercivity Is a Necessary Condition
We prove that the operator g + h has to be coercive if the equation in Theorem
6.1 is true, and thus that the coercivity of the operator g + h in Theorem 6.1 is
a necessary condition. We let (M, g) be a smooth compact Riemannian manifold
of dimension n 3, and for h C 0, (M ), (0, 1), and u H12 (M ), we let

|u|2 + hu2 dvg .
Ih (u) =
M
Then we let h = inf u Ih (u) where the infimum is with respect to the us in
H12 (M ) such that M u2 dvg = 1. The operator g + h is said to be coercive
[see (2.1.1)] if there exists C > 0 such that Ih (u) Cu2H 2 for all u H12 (M ).
1
Easy claims are as follows:
C LAIM 1. The operator g + h is coercive if and only if h > 0.
uh C 2, (M ), uh > 0 everywhere, such
C LAIM 2. Whatever h is, there
exists
2
that g uh + huh = h uh and M uh dvg = 1.
C LAIM 3. If there exists u C 2, (M ) such that u > 0 and g u + hu > 0
everywhere, then g + h is coercive.
We prove claim 1 by noting that, if h > 0, then
|u|2 dvg +
hu2 dvg
M
M

|u| dvg +
hu2 dvg + (1 )h
u2 dvg
M
M
M

|u|2 dvg +
u2 dvg
M
where > 0 is chosen such that (1 )h + h . Claim 2 easily follows from

standard minimization techniques, the maximum principle, and standard bootstrap
arguments for regularity. Finally, we prove claim 3 using claim 2 by writing that
0<
uh (g u + hu) dvg
M
= h
uuh dvg
M
so that h > 0, and we are back to claim 1.
210
APPENDIX B
Now we let (h ) be a sequence of functions in C 0, (M ), (0, 1), we let

h C 0, (M ), and let (u ), u > 0, be a sequence in C 2, (M ) such that for any
,
2 1
g u + h u = u
and such that h h in C

(1 )u0 +
where
1
C
0,
(B1)
(M ) as +. We assume that
N
Bi u (1 + )u0 + C
i=1
Bi
(B2)
i=1
n2
2
Bi (x) =
N
i,
2i,
dg (xi, ,x)2
n(n2)
the xi, s are converging sequences of points in M , the i, s are positive real
numbers such that i, 0 as +, C is a positive constant independent of
and x, N N , and u0 H12 (M ) is a nonnegative function such that u u0
in H12 (M ). Then we claim that, necessarily, the operator g + h is coercive.
We prove the claim in what follows. We let = h and u = uh . By the
maximum principle and regularity theory, u0 C 2, (M ),
g u0 + h u0 = (u0 )2
and either u0 0 or u0 > 0 everywhere. If u0 > 0 everywhere, then g + h is

coercive from claim 3. We may therefore assume that u0 0. As an easy remark,
we necessarily have that 0, since for any u H12 (M ),
Ih (u) = lim Ih (u) .
+
We proceed by contradiction and assume that g + h is not coercive. Then, from

claim 1, this means that = 0. Using claim 2 we then get that u C 2, (M ),
u > 0, is such that
g u + h u = 0 .
We let 1, be such that 1, i, for all i. From (B1) we can write that
2 1
(g u + h u ) u dvg =
u
u dvg .
M
(B3)
Since = 0,
(g u + h u ) u dvg
M
=
u (g u + h u ) u dvg +
(h h ) u u dvg
M
M
=
(h h ) u u dvg
M
and since h h in C 0, (M ), we get that

(g u + h u ) u dvg = o
M

u dvg
(B4)
211
COERCIVITY IS A NECESSARY CONDITION
Then, using (B2), we have that

u dvg C
M
Bi dvg
i=1
and we can write that
n
2 1
Bi dvg + O i,
Bxi, ()
for all i, where > 0 is small. We have that
n
2 1
Bi dvg Ci,
Bxi, ()
+1
2
Ci,
rn1
2i, + r2
Bi dvg
i,
dr
n2
2
rn1
n2
(1 + r2 ) 2

n
i,
2 +1
Ci,
rdr
C+
0
dr
so that
Bxi, ()
n
2 1
.
Bi dvg = O i,
Then, by the definition of 1, and from (B4),
n
2 1
.
(g u + h u ) u dvg = o 1,
(B5)
Still, using (B2), and since u > 0 everywhere, we can write that
2 1
2 1
u
u dvg C
u
dvg
Bx1, (1, )
Bx1, (1, )
B1
2 1
dvg .
Noting that
Bx1, (1, )

1 2 1
dvg C
1,
0
n
2 1
C1,
it follows that
1,
2
1, + r2
rn1 dr
rn1
(1 + r2 )
n
2 1
2
u
u dvg C1,
n+2
2
n+2
2
dr
(B6)
for some C > 0 independent of . Inserting (B5) and (B6) into (B3), we get the
desired contradiction. This proves the above claim.
Bibliography
[1] Adimurthi, Pacella, F., and Yadava, S.L., Characterization of concentration

points and L -estimates for solutions of a semilinear Neumann problem
involving the critical Sobolev exponent, Differential Integral Equations, 8,
4168, 1995.
[2] Allendoerfer, C. B., The Euler number of a Riemann manifold, Amer. J.
Math., 62, 243248, 1940.
[3] Allendoerfer, C. B., and Weil, A., The Gauss-Bonnet theorem for Riemannian polyhedra, Trans. Amer. Math. Soc., 53, 101129, 1943.
[4] Ambrosetti, A., and Rabinowitz, P. H., Dual variational methods in critical
point theory and applications, J. Funct. Anal., 14, 349381, 1973.
[5] Atkinson, F. V., and Peletier, L. A., Elliptic equations with nearly critical
growth, J. Diff. Equ., 70, 349365, 1987.
[6] Aubin, T., Equations

differentielles non lineaires et problème de Yamabe
concernant la courbure scalaire, J. Math. Pures Appl., 55, 269296, 1976.
[7]
, Nonlinear analysis on manifolds. Monge-Ampère equations,

Grundlehren der Mathematischen Wissenschaften, 252, Springer, New York,
1982.
[8] Avez, A., Applications de la formule de Gauss-Bonnet-Chern aux varietes a`

quatre dimensions, C. R. Acad. Sci. Paris, 256, 54885490, 1963.
[9] Berestycki, H., Nirenberg, L., and Varadhan, S.R.S, The principal eigenvalue and maximum principle for second order elliptic operators in general
domains, Comm. Pure Appl. Math., 47, 4792, 1994.
[10] Berger, M., Les varietes riemanniennes (1/4)-pincees, Ann. Scuola Norm.
Sup. Pisa 14, 161170, 1960.
[11] Brezis, H., Problèmes de convergence dans certaines EDP non lineaires
et applications geometriques, Seminaire Goulaouic-Meyer-Schwartz, Ecole
Polytechnique, 1983-84.
[12] Brezis, H., and Coron, J.M., Convergence de solutions de H-systèmes et
application aux surfaces a` courbure moyenne constante, C. R. Acad. Sci.
Paris, 298, 389392, 1984.
214
[13]
BIBLIOGRAPHY
, Convergence of solutions of H-systems or how to blow bubbles,

Arch. Rational Mech. Anal., 89, 2156, 1985.
[14] Brezis, H., and Lieb, E., A relation between pointwise convergence of functions and convergence of functionals, Proc. Amer. Math. Soc., 88, 486490,
1983.
[15] Brezis, H., and Nirenberg, L., Positive solutions of nonlinear elliptic equations involving critical Sobolev exponents, Comm. Pure Appl. Math., 36,
437477, 1983.
[16] Brezis, H., and Peletier, L. A., Asymptotics for elliptic equations involving
critical Sobolev exponents, in Partial Differential Equations and the Calculus of Variations, eds. F. Colombini, A. Marino, L. Modica, and S. Spagnalo,
Birkhauser, Basel, 1989.
[17] Caffarelli, L. A., Gidas, B., and Spruck, J., Asymptotic symmetry and local behavior of semilinear elliptic equations with critical Sobolev growth,
Comm. Pure Appl. Math., 42, 271297, 1989.
[18] Chang, A., and Yang, P., A perturbation result in prescribing scalar curvature
on S n , Duke Math. J., 64, 2769, 1991.
[19] Chang, A., Gursky, M., and Yang, P., The scalar curvature equation on 2and 3-spheres, Calc. Var. Partial Differential Equations, 1, 205229, 1993.
[20] Chavel, I., Riemannian geometryA modern introduction, Cambridge Tracts
in Mathematics, 108, Cambridge University Press, Cambridge, 1993.
[21] Chern, S. S., A simple intrinsic proof of the Gauss-Bonnet formula for closed
Riemannian manifolds, Ann. Math., 45, 747752, 1944.
[22] do Carmo, M. P., Riemannian geometry, Mathematics: Theory and Applications, Birkhauser, Boston, 1992.
[23] Druet, O., The best constants problem in Sobolev inequalities, Math. Ann.,
314, 327346, 1999.
[24]
, Optimal Sobolev inequalities and extremal functions. The threedimensional case, Indiana Univ. Math. J., 51, 6988, 2002.
[25]
, Elliptic equations with critical Sobolev exponents in dimension 3,

Ann. Inst. H. Poincare. Anal. Non Lineaire, 19, 125142, 2002.
[26]
, Sharp local isoperimetric inequalities involving the scalar curvature, Proc. Amer. Math. Soc., 130, 23512361, 2002.
[27]
, From one bubble to several bubbles: The low-dimensional case, J.

Differential Geom., 63, 399473, 2003.
BIBLIOGRAPHY
215
, Compactness for Yamabe metrics in low dimensions, Preprint,
[28]
2003.
[29] Druet, O., and Hebey, E., The AB program in geometric analysis. Sharp
Sobolev inequalities and related problems, Memoirs of the American Mathematical Society, MEMO/160/761, 2002.
[30]
, Asymptotics for sharp Sobolev-Poincare inequalities on compact

Riemannian manifolds, Adv. Differential Equations, 7, 14091478, 2002.
[31]
, Blow-up examples for second order elliptic PDEs of critical

Sobolev growth, Trans. Amer. Math. Soc., to appear.
[32] Druet, O., Hebey, E., and Robert, F., A C 0 -theory for the blow-up of second order elliptic equations of critical Sobolev growth, Electronic Research
Announcements of the AMS, 9, 1925, 2003.
[33] Druet, O., and Robert, F., Asymptotic profile for the sub-extremals of the
sharp Sobolev inequality on the sphere, Comm. P.D.E., 25, 743778, 2001.
[34] Ekeland, I., and Ghoussoub, N., Selected new aspects of the calculus of
variations in the large, Bull. Amer. Math. Soc., 39, 207265, 2002.
[35] Fenchel, W., On total curvatures of Riemannian manifolds: I, J. London
Math. Soc., 15, 1522, 1940.
[36] Gallot, S., Hulin, D., and Lafontaine, J., Riemannian geometry, second edition, Universitext, Springer, Berlin, 1993.
[37] Gilbarg, G., and Trudinger, N. S., Elliptic partial differential equations
of second order, second edition, Grundlehren der Mathematischen Wissenschaften, 224, Springer, Berlin, 1983.
[38] Giraud, G., Sur le problème de Dirichlet generalise, Ann. Sc. Ecole Norm.
Sup., 46, 131245, 1929.
[39] Hamilton, R. S., Three-manifolds with positive Ricci curvature, J. Differential Geom., 17, 255306, 1982.
[40] Han, Z. C., Asymptotic approach to singular solutions for nonlinear elliptic
equations involving critical Sobolev exponent, Ann. Inst. H. Poincare. Anal.
Non Lineaire, 8, 159174, 1991.
[41] Han, Q., and Lin, F., Elliptic partial differential equations, CIMS Lecture
Notes, Courant Institute of Mathematical Sciences, Vol. 1, 1997. Second
edition published by the American Mathematical Society, 2000.
[42] Hebey, E., Courbure scalaire et geometrie conforme, J. Geom. Phys., 10,
345380, 1993.
216
BIBLIOGRAPHY
[43]
, Introduction a` lanalyse non lineaire sur les varietes, Diderot Editeur, Fondations, Paris, 1997.
[44]
, Nonlinear analysis on manifolds: Sobolev spaces and inequalities, CIMS Lecture Notes, Courant Institute of Mathematical Sciences, Vol.
5, 1999. Second edition published by the American Mathematical Society,
2000.
[45]
, Nonlinear elliptic equations of critical Sobolev growth from a dynamical viewpoint, Contemporary Mathematics, Proceedings of the conference in honor of H. Brezis and F. Browder, to appear.
[46]
, Variational methods and elliptic equations in Riemannian geometry, Preprint, Notes from lectures given at ICTP, 2003.
[47] Hebey, E., and Vaugon, M., The best constant problem in the Sobolev embedding theorem for complete Riemannian manifolds, Duke Math. J., 79,
235279, 1995.
[48]
, Meilleures constantes dans le theorème dinclusion de Sobolev,

Ann. Inst. H. Poincare. Anal. Non Lineaire, 13, 5793, 1996.
, From best constants to critical functions, Math. Z., 237, 737767,
[49]
2001.
[50] Jost, J., Riemannian geometry and geometric analysis, Universitext,

Springer, Berlin, 1995.
[51] Klingenberg, W., Contributions to Riemannian geometry in the large, Ann.
Math., 69, 654666, 1959.
[52]
, Uber
Riemannsche Mannigfaltigkeiten mit positiver Krummung,
Comment. Math. Helv., 35, 4754, 1961.
[53] Kobayashi, S., and Nomizu, K., Foundations of differential geometry, 2

vols., Interscience, Wiley, New York, 1963, 1969.
[54] Lee, J. M., and Parker, T. H., The Yamabe problem, Bull. Amer. Math. Soc.,
17, 3791, 1987.
[55] Li, Y. Y., Prescribing scalar curvature on S n and related problems, Part I, J.
Diff. Equations, 120, 319410, 1995.
[56]
, Prescribing scalar curvature on S n and related problems, Part II:

existence and compactness, Comm. Pure Appl. Math., 49, 541597, 1996.
[57] Li, Y. Y., and Zhu, M., Yamabe type equations on three dimensional Riemannian manifolds, Communications in Contemporary Math., 1, 150, 1999.
[58] Lions, P.-L., The concentration-compactness principle in the calculus of
variations. The limit case I, II, Rev. Mat. Iberoamericana, no. 1, 145-201;
no. 2, 45121, 1985.
BIBLIOGRAPHY
217
[59] Lohkamp, J., Metrics of negative Ricci curvature, Ann. Math., 140, 655683,
1994.
[60] Obata, M., The conjectures on conformal transformations of Riemannian
manifolds, J. Differential Geom., 6, 247258, 1971/72.
[61] Rauch, H. E., A contribution to differential geometry in the large, Ann.
Math., 54, 3855, 1951.
[62] Robert, F., Asymptotic behavior of a nonlinear elliptic equation with critical
Sobolev exponent The radial case, Adv. Differential Equations, 6, 821846,
2001.
[63]
, Asymptotic behavior of a nonlinear elliptic equation with critical

Sobolev exponent The radial case II, Nonlinear Differ. Equ. Appl., 9, 361
384, 2002.
[64] Sacks, P., and Uhlenbeck, K., On the existence of minimal immersions of
2-spheres, Ann. Math., 113, 124, 1981.
[65] Sakai, T., Riemannian geometry, Translations of Mathematical Monographs,
149, American Mathematical Society, Providence, R.I., 1996.
[66] Schoen, R., Conformal deformation of a Riemannian metric to constant
scalar curvature, J. Differential Geom., 20, 479495, 1984.
[67]
, Variational theory for the total scalar curvature functional for Riemannian metrics and related topics, in Topics in calculus of variations (Montecatini Terme, 1987). Lecture Notes in Mathematics, 1365, Springer, Berlin,
1989, 120154.
[68]
, Courses at Stanford University (1988) and New York University

(1989), unpublished.
[69]
, A report on some recent progress on nonlinear problems in geometry, Surveys in Differential Geometry, Cambridge, Mass., 1990, 201241.
[70]
, On the number of constant scalar curvature metrics in a conformal

class, Differential Geometry: A symposium in honor of Manfredo Do Carmo,
eds. H. B. Lawson and K. Tenenblat, Pitman Monograph Surveys in Pure and
Applied Mathematics, 52, 311320, 1991.
[71] Schoen, R., and Zhang, D., Prescribed scalar curvature on the n-sphere,
Calc. Var. Partial Differential Equations, 4, 125, 1996.
[72] Spivak, M., A comprehensive introduction to differential geometry, 5 vols.,
Publish or Perish, Houston, Tex., 1970.
[73] Struwe, M., A global compactness result for elliptic boundary problems involving limiting nonlinearities, Math. Z., 187, 511517, 1984.
218
BIBLIOGRAPHY
[74] Trudinger, N. S., Remarks concerning the conformal deformation of Riemannian structures on compact manifolds, Ann. Scuola Norm. Sup. Pisa, 22,
265274, 1968.
[75] Tsukamoto, Y., On Riemannian manifolds with positive curvature, Mem.
Fac. Sci. Kyushu Univ. Ser. A, 15, 9096, 1961/1962.
[76] Wente, H. C., Large solutions to the volume constrained Plateau problem,
Arch. Rational Mech. Anal., 75, 5977, 1980.

Blow-Up Theory For Elliptic PDEs in Riemannian Geometry

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Blow-Up Theory For Elliptic PDEs in Riemannian Geometry

Uploaded by

Copyright:

Available Formats

Blow-up Theory for Elliptic

This page intentionally left blank