Topological Methods in Nonlinear Analysis

Journal of the Juliusz Schauder Center

Volume 14, 1999, 138
EFFECT OF THE DOMAIN GEOMETRY
ON THE EXISTENCE OF MULTIPEAK SOLUTIONS
FOR AN ELLIPTIC PROBLEM
E. Norman Dancer Shusen Yan
Abstract. In this paper, we construct multipeak solutions for a singu-
larly perturbed Dirichlet problem. Under the conditions that the distance
function d(x, ) has k isolated compact connected critical sets T
1
, . . . , T
k
satisfying d(x, ) = c
j
= const., for all x T
j
, min
i=j
d(T
i
, T
j
) >
2 max
1jk
d(T
j
, ), and the critical group of each critical set T
i
is non-
trivial, we construct a solution which has exactly one local maximum point
in a small neighbourhood of T
i
, i = 1, . . . , k.
1. Introduction
Let be a bounded domain in R
N
. Consider
(1.1)
_

2
u +u = u
p1
in ,
u > 0 in ,
u = 0 in ,
where is a small positive number, 2 < p < 2N/(N2) if N 3 and 2 < p <
if N = 2.
In the past few years, a lot of work has been done on the existence and mul-
tiplicity of the solutions for (1.1). First, Ni and Wei [27] proved that as 0,
1991 Mathematics Subject Classication. 35J65, 35B25, 58E05.
Key words and phrases. Peaks, nonlinear, small diusion, Conley index, Clarke derivative.
This research was supported by ARC.
c 1999 Juliusz Schauder Center for Nonlinear Studies
1
2 E. N. Dancer S. Yan
the least energy solution has exactly one local maximum point and this local
maximum point tends to a point which attains the global maximum of the dis-
tance function d(x, ). From then, the eect of the distance function on the
existence of single peak solution, that is, solution which has exactly one local
maximum point, has been extensively studied and single peak solutions with the
peak near various kind of critical points of the distance function have been con-
structed. See [21], [23], [28], [32], [33]. On the other hand, if the distance function
has several critical points, there are a few works concerning the construction of
multipeak solutions (see [6], [7], [18]).
There is another direction to study the multiplicity of the single peak solu-
tions and the existence of multipeak solution: the eect of the domain topology.
In [2], Benci and Cerami proved that (1.1) has at least Cat

() single peak
solution if > 0 is small enough. Later, Benci, Cerami and Pasasseo proved
in [3] that if is not contractible, the number of the solutions of (1.1) is at least
Cat

() + 1. Concerning the eect of the domain topology on the existence

of multipeak solution, we proved in [17] that if the homology of the domain is
nontrivial, then for any positive integer k, (1.1) has at least one k-peak solution
provided > 0 is small enough. The method in [17] can also be modied to
show that there is a two-peak solution for small > 0 if is not contractible.
See also [14] for an early result on the existence of two-peak solutions.
The aim of this paper is to construct multipeak solutions for (1.1) provided
that d(x, ) has several critical points whose critical groups are nontrivial. We
rst give some denitions and recall some basic results.
Let f(x) be a Lipschitz continuous function dened on R
N
. The Clarke
derivative of f is dened as follows (see [10]):
f(x) = R
N
: f
0
(x, v) , v), for all v R
N
,
where
f
0
(x, v) = lim
h0,0
+
f(x +h +v) f(x +h)

.
A point x
0
is called a critical point of f if 0 f(x
0
). Let T be an isolated
connected critical set of f in the following sense: 0 f(x), f(x) = c for each
x T, and f has no other critical point in a small neighbourhood | of T. Then
we dene the critical group of f on T as follows:
C
q
(f, T) = H
q
(f
c
|, (f
c
T) |),
where q = 0, 1, . . . , f
c
= x : f(x) c.
We stress here that all the cohomologies in this paper are with the coecients
in the same eld.
Multipeak Solutions 3
The distance function d(x, ) is Lipschitz continuous and its Clarke deriv-
ative is
(1.2) d(x, ) = co (

(x) x),
where

(x) = y : y , [y x[ = d(x, ) and co S denotes the convex

hull of the set S (see [10]).
Let U(y) be the unique positive solution (see [22]) of
_

_
u +u = u
p1
in R
N
,
u H
1
(R
N
),
u(0) = max
yR
N u(y).
It is well known that U(y) is radially symmetric about the origin, decreasing
and
lim
|y|
U(y)e
|y|
[y[
(N1)/2
= c
0
> 0.
Dene u, v)

=
_

2
Du Dv + uv, for all u, v H
1
0
(), |u|

= u, u)
1/2

.
For any z R
N
, > 0, let U
,z
(y) =: U((y z)/).
We denote by P
,
v the solution of the following problem:
_

2
u +u = [v[
p2
v in ,
u H
1
0
().
By the maximum principle, we know P
,
U
,z
> 0.
For any x
j
, j = 1, . . . , k, dene
E
,x,k
=
_
v H
1
0
() : P
,
U
,x
j
, v)

=
_
P
,
U
,x
j
x
ji
, v
_

= 0,
j = 1, . . . , k, i = 1, . . . , N
_
.
The main results of this paper are the following.
Theorem 1.1. Suppose that k 1 is an integer. Let T
j
be an isolated
critical set of the distance function d(x, ) with d(x, ) = c
j
for all x T
j
,
and D
j
be a small neighbourhood of T
j
such that d(x, ) has no critical point in
D
j
T
j
, j = 1, . . . , k. Suppose that the critical group C(d(x, ), T
j
) is nontrivial
for j = 1, . . . , k. If
(1.3) min
i=j
d(D
i
, D
j
) > 2 max
1jk
max
xD
j
d(x, ),
then there is an
0
> 0, such that for each (0,
0
], (1.1) has a solution of the
form
(1.4) u

=
k

j=1

,j
P
,
U
,x
,j
+v

,
4 E. N. Dancer S. Yan
where v

E
,x

,k
, and as 0,
(1.5)
,j
1, x
,j
x
j
T
j
, |v

= o(
N/2
), j = 1, . . . , k.
Usually, we call the solution of the form (1.4) satisfying (1.5) a k-peak solu-
tion. Apart from the above existence result, we have the following nonexistence
result.
Theorem 1.2. Suppose that is strictly convex. Then for each integer
k 2, there is an
0
> 0, such that for each (0,
0
], (1.1) does not have any
k-peak solution.
Remark 1.3. It can be proved that a domain is strictly convex if and
only if for each point x , T
x
( x) = , where T
x
is the tangent plane
of at x.
Remark 1.4. In [33], Wei proved Theorem 1.2 for the case k = 2 and asked
whether the conclusion was still true for k 3.
Using (1.2), we can check that the peak of any single peak solution must
converge to a critical point of the distance function as 0. See the discussion
in the beginning of Section 3. So it is natural to ask what kind of critical point of
the distance function can generate a single peak solution with its peak near this
critical point. The example given in Section 4 shows that some of the critical
points of the distance function may not generate a single peak solution. On the
other hand, for an isolated strict local maximum point, or a critical point such
that the distance function is dierentiable on the boundary of a small neighbour-
hood of this point and the degree of Dd(x, ) in this small neighbourhood is
not zero, then there is a single peak solution with its peak nearby (see [6], [18],
[21], [23], [28], [32], [34]). In Section 3, we prove some results on the nontriviality
of the critical group. These results enable us to conclude that the critical groups
of local maximum points, or some kind of saddle points, or critical point with
nonzero degree are nontrivial. We also give in Section 3 an example where the
isolated critical point of the distance function is not the kind of saddle point
dened in [9], and the distance function is not dierentiable on the boundary of
any small neighbourhood of this point, but whose corresponding critical group
is nontrivial. This example also shows that the structure of d(x, ) in a small
neighbourhood of an isolated critical point may be very complicated even if its
corresponding critical group is nontrivial.
As we shall see, to construct a single peak solution for (1.1) with its peak
near a designated critical point x
0
of d(x, ) essentially reduces to nd a con-
dition under which x
0
is stable subject to suitable small perturbations (see [23]
for the precise denition). Thus we need to discuss the stability problem for the
critical point of a function f(x). In order to make it possible to glue together
Multipeak Solutions 5
the single peak solutions, we also need to study whether (x
1
, x
2
) is a stable crit-
ical point of f
1
(x
1
) + f(x
2
) if x
i
is a stable critical point of f
i
(x
i
), i = 1, 2. In
the smooth case, we know that if the corresponding critical group is nontrivial,
then the critical point is stable. Moreover, C(f
1
+f
2
, (x
1
, x
2
)) is nontrivial and
thus (x
1
, x
2
) is a stable critical point of f
1
(x
1
) +f(x
2
), if C(f
i
, x
i
) is nontrivial,
i = 1, 2. In the smooth case, it is also well known that the corresponding critical
groups of strict local maximum points, strict local minimum points, nondegen-
erate critical points, critical points with nonzero degree are nontrivial. For the
distance function, we know it has no minimum point inside . Wei in [34] gave
the denition for a critical point of d(x, ) to be nondegenerate. However, we
prove in Section 4 that under Weis denition, a nondegenerate critical point
of d(x, ) must be a strict local maximum point. On the other hand, the dis-
advantage to use classical degree theory to deal with the stability problem is
that in many cases, it may not be possible to nd a small neighbourhood for the
isolated critical point such that d(x, ) is dierentiable on the boundary of this
neighbourhood and d(x, ) has no other critical point in this neighbourhood.
For these reasons, to deal with the stability problem for a critical point of the
distance function, which is not a local maximum point, we need to generalize
the degree theory and the critical group theory to the nonsmooth problems. Our
result in Section 3 shows that as in the smooth case, the nonzero degree implies
the nontriviality of the critical groups. This is one of the reasons we use critical
groups to deal with the stability problem for the critical points. There is another
advantage to use the critical groups in the nonsmooth problems. To calculate
the degree of the Clarke derivative of a Lipschitz continuous function, one needs
to choose a smooth approximating vector eld and calculate the degree of this
smooth vector eld, while the calculation of the critical groups does not involve
the choice of a smooth approximating vector eld. So in practice, the critical
group is calculable. In addition, we can also construct an example of a critical
set which has zero degree but whose critical group is non-zero. See Section 3.
In [21], Grossi and Pistoia gave the denition for a critical value to be topo-
logically nontrivial. But it is not clear whether c
1
+ c
2
is a critical value of
f
1
(x
1
) +f
2
(x
2
) topologically nontrivial if c
i
is a critical value of f
i
(x
i
), i = 1, 2,
which is topologically nontrivial. Thus it seems dicult to glue together the sin-
gle peak solutions constructed in [21]. Another disadvantage of using this idea
to construct single peak solution with its peak near a designated critical point
is that extra conditions on this critical point are needed.
The basic idea to construct multipeak solutions with their peaks close to some
specic critical points is to glue some single peak solutions together. To make
this idea work, usually it is required that dierent critical points be suitably
separated. In [6], [7], [18], condition (1.3) is imposed so the contribution from
6 E. N. Dancer S. Yan
the interaction between dierent peaks is negligible. It seems that (1.3) is almost
necessary for Theorem 1.1 to hold, at least this is true if d(P
j
, ) = d(P
i
, ),
i ,= j. In [6], [7], reduction methods were used to construct k-peak solutions
with the peaks near the local maximum points or saddle points (in the strong
sense) P
1
, . . . , P
k
of d(x, ). But in these two papers, it is required that
(1.6) d(P
j
, ) = d(P
i
, ), i, j = 1, . . . , k.
In [18], del Pino, Felmer and Wei used a variational method to construct a k-
peak solution with its peaks close to some local maximum points P
1
, . . . , P
k
without assuming (1.6). But it seems that the variational method is not easy
to use to construct multipeak solutions with the peaks near some critical points
which are not local maximum points. In this paper, we still use the reduction
method to construct multipeak solution for (1.1). Without condition (1.6), the
rst small order term of the energy of the single peak solution with its peak
near P
i
is of dierent order from that of the single peak solution with its peak
near P
j
. To glue together single peak solutions involving dierent small order
terms needs a lot of work. The new idea here is to split the small part v

into k
dierent part, each of which is very close to the corresponding small part of
a single peak solution. By this result, we are able to prove that the reduced
problem is just a small perturbation of that corresponding to a single peak
solution, provided (1.3) holds. In Section 4, we will give an example showing
that the nontriviality of the critical group is almost necessary to obtain the result
of Theorem 1.1. Theorem 1.1 is new even if all the critical points are saddle points
in the strong sense. The nontriviality of C(d, x
i
), i = 1, . . . , k, is the weakest
condition we know to guarantee the existence of a k-peak solution which has
exactly one peak near each x
i
. Our methods have a number of advantages over
other work. Firstly, it applies to critical points which are not saddle points in
the sense of Rabinowitz. Moreover, unlike a number of other recent works [2],
[14], [17], we can also control the location of the peaks of the solutions.
Note that to obtain our results, we construct the Conley index, critical groups
and degree of an isolated critical point of a locally Lipschitz function. This seems
of independent interest.
The method to prove Theorem 1.1 can be used to obtain the same result
for the Dirichlet problem in exterior domains under the condition that the dis-
tance function has some critical points with nontrivial critical group. Unlike the
bounded domain case, the distance function on exterior domain does not always
have a critical point. However we proved in [16] that the Dirichlet problem in
an exterior domain always has a two peak solution and if the domain is the
complement of a bounded convex set, then the peaks of any two peak solution
Multipeak Solutions 7
move to innity. For other results on the Dirichlet problem in exterior domains,
the readers can refer to [4], [5], [26].
We can also use this technique to glue together a boundary peak solution
and an interior peak solution for the singularly perturbed Neumann problem.
This is discussed briey in Section 5.
This paper is organised as follows. In Section 2, we reduce the problem
of nding a multipeak solution for (1.1) to a nite dimensional problem. The
main ingredient of Section 2 is to split the small term into dierent parts. This
technique is important in gluing dierent single peak solutions involving dier-
ent order small terms. Section 3 contains the discussion of the Conley index
of the generalized gradient of a Lipschitz function and its critical group. Theo-
rems 1.1 and 1.2 are proved in Section 4. Some basic estimates are presented in
the appendix.
2. Reduction of the problem
Let
(2.1) I(u) =
1
2
_

(
2
[Du[
2
+u
2
)
1
p
_

[u[
p
, u H
1
0
().
For xed integer k > 0, let
(k) = (
1
, . . . ,
k
) R
k
, (2.2)
x = (x
1
, . . . , x
k
) R
kN
, x
i
R
N
, i = 1, . . . , k. (2.3)
Dene D
k,
= x : x
i
, d(x
i
, ) , i = 1, . . . , k, [x
i
x
j
[ 2, i ,= j,
and M
,
= ((k), x, v) : [
i
1[ , i = 1, . . . , k; x D
k,
, v E
,x,k
, |v|

N/2
. Let
(2.4) J((k), x, v) = I
_
k

i=1

i
P
,
U
,x
i
+v
_
, (, x, v) M
,
.
It is well known that if > 0 is small enough, ((k), x, v) M
,
is a critical
point of J((k), x, v) if and only if u =

k
i=1

i
P
,
U
,x
i
+v is a positive critical
point of I(u) (see [30]). So we just need to nd ((k), x, v) M
,
and A
l
, B
li
,
l = 1, . . . , k, i = 1, . . . , N, such that
(2.5)
J((k), x, v)
x
li
=
N

j=1
B
lj
_

2
P
,
U
,x
l
x
li
x
lj
, v
_

, i = 1, . . . , N, l = 1, . . . , k,
(2.6)
J((k), x, v)

l
= 0, l = 1, . . . , k,
(2.7)
J((k), x, v)
v
=
k

l=1
A
l
P
,
U
,x
l
+
k

l=1
N

j=1
B
lj
P
,
U
,x
l
x
lj
.
8 E. N. Dancer S. Yan
The aim of this section is to reduce the problem of nding a critical point for
J((k), x, v) to that of nding a critical point for a function dened in a nite
dimensional domain. The proof of the existence of ((k), v) satisfying (2.6)
and (2.7) for each xed x is quite standard. The crucial part of this section is
to split the small term v into dierent parts.
Proposition 2.1. There are
0
> 0 and > 0, such that for each (0,
0
],
there is a unique C
1
-map (

(k, x), v
,k
(x)) : D
k,
R
k
H
1
0
(), satisfying
v
,k
(x) E
,x,k
, (2.6) and (2.7). Besides, if k 2 and l = 1, . . . , k, then
(2.8) [
,l
1[ = O
_
k

j=1
e
(1+)d(x
j
,)/
+

i=j
e
(1+)|x
i
x
j
|/2
_
,
|v
,k
|

= O
_

N/2
_
k

j=1
e
(1+)d(x
j
,)/
+

i=j
e
(1+)|x
i
x
j
|/2
__
, (2.9)
and if k = 1, then
(2.10) [
,l
1[
N/2
+|v
,1
|

= O(
N/2
e
(1+)d(x,)/
),
where is some positive constant. Moreover, for l = 1, . . . , k, i = 1, . . . , N,
we have
(2.11) A
l
, B
li
= O
_

j=1
e
(2)d(x
j
,)/
+

h=j
U
_
[x
h
x
j
[

__
.
Proof. The proof of the existence part is standard (see [6], [15], and also
[1], [30]). The estimates (2.8) and (2.9) or (2.10) follow from the same procedure
as in Proposition 2.3 of [15] and Lemmas A.l and A.2. Finally, we can solve
an appropriate system as in [30, pp. 2223], to get (2.11). We thus omit the
details.
We also need the following pointwise estimate for v
,k
(x, y).
Lemma 2.2. Let v
,k
(x, y) be the map obtained in Proposition 2.1. For any
small > 0, there is a > 0 such that
v
,k
(x, y) = O
_
e
/
k

j=1
U
1
_
[y x
j
[

__
.
Multipeak Solutions 9
Proof. For xed x, let w(y) = v
,k
(x, y). By (2.7), we know that v
,k
(x, y)
satises
_
J
v
,
_
=
k

j=1
A
j
P

U
,x
j
, ) +
k

j=1
N

i=1
B
ij
_
P

U
,x
j
x
ji
,
_
=
k

j=1
A
j
_

U
p1
,x
j
+
k

j=1
N

i=1
(p 1)B
ij
_

U
p2
,x
j
U
,x
j
x
ji
.
Thus, w satises
w +w =
_
k

j=1

j
P

U
x
j
/
+w
_
p1

j=1

j
U
p1
x
j
/
+
k

j=1
A
j
U
p1
x
j
/
+
k

j=1
N

i=1
(p 1)B
ij
U
p2
x
j
/
U
x
j
/
x
ij
=: F(y),
y

=: y : y ; w = 0, y

. Moreover, we have the following

estimate
|w| = O
_
k

j=1
e
(1+)d(x
j
,)/
+

i=j
e
(1+)|x
i
x
j
|/2
_
,
and
[F(y)[ C(1 +[w[
p2
)[w[ +[g(y)[,
where
g(y) =C
_
k

j=1
[P

U
x
j
/
U
x
j
/
[ +

j=i
U
(p1)/2
x
i
/
U
(p1)/2
x
j
/
+
k

j=1
A
j
U
p1
x
j
/
+
k

j=1
N

i=1
(p 1)B
ij
U
p2
x
j
/
U
x
j
/
x
ij
_
.
Choose q > N/2 satisfying q(p 2) < 2N/(N 2). Thus, for any z

,
we have
_
B
1
(z)
(1 +[w[
p2
)
q
C

,
and
[g[
L
q
(B
1
(z))
C
k

j=1
[P

U
x
j
/
U
x
j
/
[
L
q
(B
1
(z))
+C
_
k

j=1
e
(2)d(x
j
,)/
+

h=j
U
_
[x
h
x
j
[

__
C
k

j=1
e
d(x
j
,)/
+C

i=j
e
(1+)|x
i
x
j
|/2
.
10 E. N. Dancer S. Yan
So it follows from Theorem 8.17 in [20] that
[w(y)[ C[w[
L
2N/(N2)
(B
1
(z))
+C[g[
L
q
(B
1
(z))
C
k

j=1
e
d(x
j
,)/
+C

i=j
e
(1+)|x
i
x
j
|/2
, for all y B
1/2
(z).
Especially, we see
[v
,k
(x, y)[ = O
_
e
l/
k

j=1
U
1
_
[y x
j
[

__
, y
k
_
j=1
B
l
(x
j
),
where l = min(d(x
j
, )/4, [x
i
x
j
[/8).
On the other hand, for > 0 small, and y

k
j=1
B
l
(x
j
), we have
[P
,
U
,x
j
(y)[ Ce
|yx
j
|/
= o(1). As a result,
F
_
y

_
= o(1)v
,k
+O
_
k

j=1
e
(p1)|yx
j
|/
_
,
where o(1) 0 as 0.
Let =

k
j=1
e
(1/2)|yx
j
|/
. By direct calculation, we see
2
+
c

, where c

> 0 is a constant depending on . Let = v

,k
. Since

2
v
,k
+v
,k
= F
_
y

_
= o(1)v
,k
+O
_
k

j=1
e
(p1)|yx
j
|/
_
for any y

k
j=1
B
l
(x
j
), we have

2
+ (1 o(1)) =
2
+ (1 o(1)) O
_
k

j=1
e
(p1)|yx
j
|/
_
c
0
k

j=1
e
(1/2)|yx
j
|/
O
_
k

j=1
e
(p1)|yx
j
|/
_
> 0,
where c
0
> 0 is a constant depending on . Moreover, for any y or
y B
l
(x
j
), [v
,k
(y)[ (y). So it follows from the comparison theorem that
[v
,k
(x, y)[ (y), y
k
_
j=1
B
l
(x
j
),
and hence the result.
For each k 2 and x = (x
1
, . . . , x
k
), by Proposition 2.1, we can deter-
mine a map (

(k, x), v
,k
(x)) satisfying (2.6) and (2.7). Let 1 m k,
x

= (x
1
, . . . , x
m
) and x

= (x
m+1
, . . . , x
k
). Our next result shows that as
min
1im<m+1jk
[x
i
x
j
[ becomes large, (

(k, x), v
,k
(x)) will eventually
split into two parts. To be more precisely, we have
Multipeak Solutions 11
Proposition 2.3. For each k 2 and x = (x
1
, . . . , x
k
), the map
(

(k, x), v
,k
(x))
obtained in Proposition 2.1 satises
m

j=1
[
,j
(k, x)
,j
(m, x

)[ +
k

j=m+1
[
,j
(k, x)
,j
(k m, x

)[
= O(U
(1+)/2
(/)),
|v
,k
(x) v
,m
(x

) v
,km
(x

)|

= O(
N/2
U
(1+)/2
(/)),
where = min
1im<m+1jk
[x
i
x
j
[.
Proof. By (2.6), for any 1 i m, we have
(2.12)
_
k

j=1

,j
(k, x)P
,
U
,x
j
, P
,
U
,x
i
_

_
k

j=1

,j
(k, x)P
,
U
,x
j
+v
,k
(x)
_
p1
P
,
U
,x
i
= 0
and
(2.13)
_
m

j=1

,j
(m, x

)P
,
U
,x
j
, P
,
U
,x
i
_

_
m

j=1

,j
(m, x

)P
,
U
,x
j
+v
,m
(x

)
_
p1
P
,
U
,x
i
= 0.
Let
G(x) =
m

j=1

,j
(m, x

)P
,
U
,x
j
+v
,m
(x

) +v
,km
(x

).
Since
P
,
U
,x
j
, P
,
U
,x
i
)

= O(
N
U(/))
and (by Lemma 2.2)
_

v
p1
,km
(x

)P
,
U
,x
i
= O(
N
U(/))
for j > m and 1 i m, we have
(2.14)
_
m

j=1
(
,j
(k, x)
,j
(m, x

))P
,
U
,x
j
, P
,
U
,x
i
_

+O(
N
U(/))
=
_

_
k

j=1

,j
(k, x)P
,
U
,x
j
+v
,k
(x)
_
p1
P
,
U
,x
i

_
m

j=1

,j
(m, x

)P
,
U
,x
j
+v
,m
(x

)
_
p1
P
,
U
,x
i
12 E. N. Dancer S. Yan
=
_

_
m

j=1

,j
(k, x)P
,
U
,x
j
+v
,k
(x)
_
p1
P
,
U
,x
i

G
p1
(x)P
,
U
,x
i
+O(
N
U(/))
=(p 1)
_

G
p2
(x)
m

j=1
(
,j
(k, x)
,j
(m, x

))P
,
U
,x
j
P
,
U
,x
i
+ (p 1)
_

G
p2
(x)(v
,k
(x) v
,m
(x

) v
,km
(x

))P
,
U
,x
i
+O
_

N
m

j=1
[
,j
(k, x)
,j
(m, x

)[
1+
_
+O(|v
,k
(x) v
,m
(x

) v
,km
(x

)|
1+

) +O(
N
U(/)),
for some > 0. Noting that
|P
,
U
,x
i
|
2

(p 1)
_

(P
,
U
,x
i
)
p
= (p 2)
N
(A+o(1)),
P
,
U
,x
j
, P
,
U
,x
i
)

= o(1)
and
_

G
p2
(x)P
,
U
,x
i
P
,
U
,x
j
= o(1) for i ,= j,
where o(1) 0 as 0, we can solve (2.14) to get
(2.15) [
,j
(k, x)
,j
(m, x

)[
= O(U(/) +
N/2
|v
,k
(x) v
,m
(x

) v
,km
(x

)|

).
Similarly,
(2.16) [
,j
(k, x)
,j
(k m, x

)[
= O(U(/) +
N/2
|v
,k
(x) v
,m
(x

) v
,km
(x

)|

).
On the other hand, by (2.7), we have
(2.17)
_
k

j=1

,j
(k, x)P
,
U
,x
j
+v
,k
(x),
_

_
k

j=1

,j
(k, x)P
,
U
,x
j
+v
,k
(x)
_
p1
= 0, E
,x,k
,
(2.18)
_
m

j=1

,j
(m, x

)P
,
U
,x
j
+v
,m
(x

),
_

_
m

j=1

,j
(m, x

)P
,
U
,x
j
+v
,m
(x

)
_
p1
= 0,
Multipeak Solutions 13
for E
,x

,m
, and
(2.19)
_
k

j=m+1

,j
(k m, x

)P
,
U
,x
j
+v
,km
(x

),
_

_
k

j=m+1

,j
(k m, x

)P
,
U
,x
j
+v
,km
(x

)
_
p1
= 0,
for E
,x

,km
. Thus,
(2.20) v
,k
(x) v
,m
(x

) v
,km
(x

), )

=
_

_
k

j=1

,j
(k, x)P
,
U
,x
j
+v
,k
(x)
_
p1

_
m

j=1

,j
(m, x

)P
,
U
,x
j
+v
,m
(x

)
_
p1

_
k

j=m+1

,j
(k m, x

)P
,
U
,x
j
+v
,km
(x

)
_
p1
,
for all E
,x,k
. Let
(2.21) G
1
(x) =
m

j=1

,j
(m, x

)P
,
U
,x
j
+v
,m
(x

)
+
k

j=m+1

,j
(k m, x

)P
,
U
,x
j
+v
,km
(x

).
Then, by Lemma 2.2,
(2.22)
_

G
p1
1
(x) =
_

_
m

j=1

,j
(m, x

)P
,
U
,x
j
+v
,m
(x

)
_
p1

+
_

_
k

j=m+1

,j
(k m, x

)PU
,x
j
+v
,km
(x

)
_
p1

+O(U
(p1)/2
(/))
N/2
||

.
Combining (2.20) and (2.22), we obtain
(2.23) v
,k
(x) v
,m
(x

) v
,km
(x

), )

=
_

__
k

j=1

,j
(k, x)P
,
U
,x
j
+v
,k
(x)
_
p1
G
p1
1
(x)
_

+O(U
(p1)/2
(/))
N/2
||

=(p 1)
_

G
p2
1
(x)(v
,k
(x) v
,m
(x

) v
,km
(x

))
14 E. N. Dancer S. Yan
+ (p 1)
_

G
p2
1
(x)
m

j=1
(
,j
(k, x)
,j
(m, x

))P
,
U
,x
j

+ (p 1)
_

G
p2
1
(x)
k

j=m+1
(
,j
(k, x)
,j
(k m, x

))P
,
U
,x
j

+
N/2
O
_
m

j=1
[
,j
(k, x)
,j
(m, x

)[
1+
+
k

j=m+1
[
,j
(k, x)
,j
(mk, x

)[
1+
_
||

+O(|v
,k
(x) v
,m
(x

) v
,km
(x

)|
1+

)||

+O(U
(p1)/2
(/))
N/2
||

.
Since for any E
,x,k
, we have
(2.24)
_

G
p2
1
(x)P
,
U
,x
j
=
_

(G
p2
1
(x) (P
,
U
,x
j
)
p2
)P
,
U
,x
j

+
_
((P
,
U
,x
j
)
p1
U
p1
,x
j
) = o(1)
N/2
||

,
for j = 1, . . . , k. Consequently, from (2.23), (2.15), (2.16) and (2.24), we obtain
(2.25) v
,k
(x) v
,m
(x

) v
,km
(x

), )

=(p 1)
_

G
p2
1
(x)(v
,k
(x) v
,m
(x

) v
,km
(x

))
+o(|v
,k
(x) v
,m
(x

) v
,km
(x

))|

||

)
+O(
N/2
U
(p1)/2
(/)||

), for all E
,x,k
.
Choose
j
and
ij
such that
=v
,k
(x) v
,m
(x

) v
,km
(x

)
+
k

j=1

j
P
,
U
,x
j
+
k

j=1
N

i=1

ij
P
,
U
,x
j
x
ij
E
,x,k
.
Noting that
v
,m
(x

), P
,
U
,x
j
),
_
v
,m
(x

),
P
,U
,x
j
x
ij
_
= O(U(/))
for j m+ 1, we obtain

j
,
ij
= O(U(/)).
So we nd from (2.25) that
||
2

CU
p1
(/)
and the result follows.
Multipeak Solutions 15
As a direct consequence of Proposition (2.3), we have the following expansion:
Proposition 2.4. Let (

(k, x), v
,k
(x)) be the map obtained in Proposi-
tion 2.3. Then for each xed i, we have
(2.26)
J
x
il
=(p 1)
_

U
p2
,x
i
U
,x
i
x
il

,x
i
(p 1)

j=i
_

U
p2
,x
i
P
,
U
,x
j
U
,x
i
x
il
+O(
N1
e
(2+)d(x
i
,)/
) +O
_

N1

j=i
U
1+
_
[x
i
x
j
[

_
+
N1

j=i
e
d(x
j
,)/
U
_
[x
i
x
j
[

__
.
(2.27)
_
J
v
, PU
,x
i
_
= O
_

N
e
(2)d(x
i
,)/
+
N

j=i
U
_
[x
i
x
j
[

__
.
Proof. We assume i = 1. Let H(y) =

k
j=1

,j
(k, x)P
,
U
,x
j
. Then
(2.28)
J
x
1l
=
_

j=1

,j
(k, x)U
p1
,x
j
P
,
U
,x
1
x
1l

_
H(y) +v
,k
(x)
_
p1
P
,
U
,x
1
x
1l
=
_

_
(H(y) +v
,k
(x))
p1
H
p1
(y)
(p 1)H
p2
(y)v
,k
(x)
_
P
,
U
,x
1
x
1l
(p 1)
_

_
H
p2
(y) (
,1
(k, x)P
,
U
,x
1
)
p2
_
v
,k
(x)
P
,
U
,x
1
x
1l
(p 1)
_

,1
(k, x)P
,
U
,x
1
_
p2
P
,
U
,x
1
x
1l
v
,k
(x)

_
H
p1
(y)
k

j=1

,j
(k, x)U
p1
,x
j
_
P
,
U
,x
1
x
1l
=: I
1
+I
2
+I
3
+I
4
,
where I
i
, i = 1, 2, 3, 4, is the natural splitting of the last formula. Denote
= min
j2
[x
j
x
1
[. We have
16 E. N. Dancer S. Yan
I
1
=
_
B
/2
(x
1
)
_
(H(y) +v
,k
(x))
p1
H
p1
(y) (2.29)
(p 1)H
p2
(y)v
,k
(x)
_
P
,
U
,x
1
x
1l

_
\B
/2
(x
1
)
_
(H(y) +v
,k
(x))
p1
H
p1
(y)
(p 1)H
p2
(y)v
,k
(x)
_
P
,
U
,x
1
x
1l
=: I
11
+I
12
.
By Lemmas A.2, 2.2 and Proposition 2.3, we have
(2.30) [I
11
[ C
1
_
B
/2
(x
1
)
[v
,k
(x)[
2
U
p2
,x
1
C
1
_
B
/2
(x
1
)
[v
,1
(x
1
)[
2
U
p2
,x
1
+C
1
_
B
/2
(x
1
)
[v
,k1
(x

)[
2
U
p2
,x
1
+C
1
_
B
/2
(x
1
)
[v
,k
(x) v
,1
(x
1
) v
,k1
(x

)[
2
U
p2
,x
1
C
N1
e
(2+)d(x
1
,)/
+C
N1
U
1+
(/).
Also, by Lemma 2.2 and Proposition 2.3, we have
(2.31) [I
12
[
1
_
\B
/2
(x
1
)
[v
,k
(x)[
p1
U
,x
1
C
1
_
\B
/2
(x
1
)
[v
,1
(x
1
)[
p1
U
,x
1
+C
1
_
\B
/2
(x
1
)
[v
,k1
(x

)[
p1
U
,x
1
+C
1
_
\B
/2
(x
1
)
[v
,k
(x) v
,1
(x
1
) v
,k1
(x

)[
p1
U
,x
1
C
N1
U
1+
(/).
Combining (2.29), (2.30) and (2.31), we obtain
(2.32) [I
1
[ C
N1
e
(2+)d(x
1
,)/
+C
N1
k

j=2
U
1+
(/).
Now we estimate I
2
.
(2.33) [I
2
[ C
1
_
B
/2
(x
1
)

H
p2
(y) (
,1
(k, x)P
,
U
,x
1
)
p2

[v
,k
(x)[P
,
U
,x
1
Multipeak Solutions 17
+C
1
_
\B
/2
(x
1
)

H
p2
(y) (
,1
(k, x)P
,
U
,x
1
)
p2

[v
,k
(x)[P
,
U
,x
1
C
1
_
B
/2
(x
1
)
k

j=2
U
p2
,x
1
U
,x
j
[v
,k
(x)[
+C
1
_
\B
/2
(x
1
)
_
k

j=2
U
,x
j
_
p2
[v
,k
(x)[U
,x
1
=: I
21
+I
22
.
But
(2.34) I
21
C
1
_
B
/2
(x
1
)
k

j=2
U
p2
,x
1
U
,x
j
[v
,k
(x) v
,1
(x
1
) v
,k1
(x

)[
+C
1
_
B
/2
(x
1
)
k

j=2
U
p2
,x
1
U
,x
j
[v
,1
(x
1
)[
+C
1
_
B
/2
(x
1
)
k

j=2
U
p2
,x
1
U
,x
j
[v
,k1
(x

)[
C
1+N/2
|v
,k
(x) v
,1
(x
1
) v
,k1
(x

)|

e
/2
+[v
,1
(x
1
)[

_
B
/2
(x
1
)
k

j=2
U
p1
,x
1
U
,x
j
+C
N1
U
1+
(/)
=O(
N1
e
/
U(/)).
Similarly,
(2.35) I
22
C
1
_
\B
/2
(x
1
)
_
k

j=2
U
,x
j
_
p2
[v
,k
(x) v
,1
(x
1
) v
,k1
(x

)[U
,x
1
+C
1
_
\B
/2
(x
1
)
_
k

j=2
U
,x
j
_
p2
[v
,1
(x
1
)[U
,x
1
+C
1
_
\B
/2
(x
1
)
_
k

j=2
U
,x
j
_
p2
[v
,k1
(x

)[U
,x
1
=O(
N1
e
/
U(/)).
Combining (2.33), (2.34) and (2.35), we obtain
(2.36) I
2
= O(
N1
e
/
U(/)).
As for the estimate of I
3
, by Proposition 2.1, Lemma 2.2 and Proposition 2.3,
we have
18 E. N. Dancer S. Yan
(2.37) I
3
= (p 1)
,1
(k, x)
_

(P
,
U
,x
1
)
p2

,1
x
1l
v
,k
(x)
(p 1)
,1
(k, x)
_

_
(P
,
U
,x
1
)
p2
U
p2
,x
1
_
U
,x
1
x
1l
v
,k
=O
_

1
_

U
p2
,x
1
[
,x
1
[[v
,k
(x)[
_
=O
_

1
_

U
p2
,x
1
[
,x
1
[[v
,1
(x
1
)[ +
1
_

U
p2
,x
1
[
,x
1
[[v
,k1
(x

)[
+
1
_

U
p2
,x
1
[
,x
1
[[v
,k
(x) v
,1
(x
1
) v
,k1
(x

)[
_
=O(
N1
e
(2+)d(x
1
,)/
) +O
_

1
[
,x
1
[

j=2
_

U
p1
,x
1
U
,x
j
_
+O
_

1
_

U
p(p2)/(p1)
,x
1
[
,x
1
[
p/(p1)
_
+O
_

1
_

[v
,k
(x) v
,1
(x
1
) v
,k1
(x

)[
p
_
=O(
N1
e
(2+)d(x
1
,)/
) +O(e
/

N1
U(/)).
Finally, we estimate I
4
.
I
4
=
_

_
H
p1
(y)
_
k

j=2

,j
(k, x)P
,
U
,x
j
_
p1
(2.38)

,1
(k, x)U
p1
,x
1
_
P
,
U
,x
1
x
1l
+
_

__
k

j=2

,j
(k, x)P
,
U
,x
j
_
p1

j=2

,j
(k, x)U
p1
,x
j
_
P
,
U
,x
1
x
1l
=: I
41
+I
42
.
We have
(2.39) I
41
=
_
B
/2(x
1
)
_
H
p1
(y)
_
k

j=2

,j
(k, x)P
,
U
,x
j
_
p1

,1
(k, x)U
p1
,x
1
_
P
,
U
,x
1
x
1l
+
_
\B
/2(x
1
)
_
H
p1
(y)
_
k

j=2

,j
(k, x)P
,
U
,x
j
_
p1

,1
(k, x)U
p1
,x
1
_
P
,
U
,x
1
x
1l
=: I
43
+I
44
.
Multipeak Solutions 19
But
(2.40) I
44
= O
__
\B
/2(x
1
)
__
k

j=2

,j
(k, x)PU
,x
j
_
p2
P
,
U
,x
1
+U
p1
,x
1
_

P
,
U
,x
1
x
1l

_
= O(
N1
U
1+
(/)).
By Lemma A.6, we see
(2.41) I
43
=
_
B
/2(x
1
)
(H
p1
(y)
,1
(k, x)U
p1
,x
1
)
P
,
U
,x
1
x
1l
+O(
N1
U
1+
(/))
=
_
B
/2(x
1
)
__

,1
(k, x)P
,
U
,x
1
_
p1

,1
(k, x)U
p1
,x
1
_
P
,
U
,x
1
x
1l
+
_
B
/2(x
1
)
(
,1
(k, x)P
,
U
,x
1
)
p2

j=2

,j
(k, x)P
,
U
,x
j
P
,
U
,x
1
x
1l
+O(
N1
U
1+
(/))
=(p 1)
_

U
p2
,x
1

,x
1
P
,
U
,x
1
x
1l
+
_

(P
,
U
,x
1
)
p2
k

j=2
P
,
U
,x
j
P
,
U
,x
1
x
1l
+O(
N1
e
(2+)d(x
1
,)/
+
N1
U
1+
(/))
=(p 1)
_

U
p2
,x
1

,x
1
U
,x
1
x
1l
+
_

U
p2
,x
1
k

j=2
P
,
U
,x
j
U
,x
1
x
1l
+O
_

N1
e
(2+)d(x
1
,)/
+
N1
U
1+
(/)
+
N1
k

j=2
e
d(x
j
,)/
U(/)
_
.
Combining (2.39), (2.40) and (2.41), we obtain
(2.42) I
41
= (p 1)
_

U
p2
,x
1

,x
1
U
,x
1
x
1l
+
_

U
p2
,x
1
k

j=2
P
,
U
,x
j
U
,x
1
x
1l
+O
_

N1
e
(2+)d(x
1
,)/
+
N1
U
1+
(/)
+
N1
k

j=2
e
d(x
j
,)/
U(/)
_
.
20 E. N. Dancer S. Yan
On the other hand,
I
42
=O
__

2i<jk
U
(p1)/2
,x
i
U
(p1)/2
,x
j

U
,x
1
x
1l

_
(2.43)
+O
__

j=2
U
p2
,x
j

,x
j

U
,x
1
x
1l

_
=O(
N1
e
|x
i
x
1
|/2|x
j
x
1
|/2(p2)|x
j
x
i
|/
)
+O
_

N1
k

j=2
e
d(x
j
,)/
U(/)
_
=O
_

N1
e
(2+)d(x
1
,)/
+
N1
U
1+
(/)
+
N1
k

j=2
e
d(x
j
,)/
U(/)
_
.
Combining (2.38), (2.42) and (2.43), we obtain
(2.44) I
4
= (p 1)
_

U
p2
,x
1

,x
1
U
,x
1
x
1l
+
_

U
p2
,x
1
k

j=2
P
,
U
,x
j
U
,x
1
x
1l
+O
_

N1
e
(2+)d(x
1
,)/
+
N1
U
1+
(/)
+
N1
k

j=2
e
d(x
j
,)/
U(/)
_
.
So (2.26) follows from (2.32), (2.36), (2.37) and (2.44). We can prove (2.27) in
a similar way.
Lemma 2.5. Let A
l
, B
lj
be the constants in Proposition 2.1. Then for each l,
we have
A
l
, B
lj
= O
_
e
(2)d(x
l
,)/
+

i=l
U
_
[x
i
x
l
[

__
.
Proof. Without loss of generality, we assume l = 1. We have
(2.45) A
1
P
,
U
,x
1
, P
,
U
,x
1
) +
N

j=1
B
1j
_
P
,
U
,x
1
x
1j
, P
,
U
,x
1
_
=
_
J
v
, P
,
U
,x
1
_
+O
_
k

l=2
U
_
[x
1
x
l
[

__
.
Multipeak Solutions 21
(2.46) A
1
_
P
,
U
,x
1
,
P
,
U
,x
1
x
1i
_
+
N

j=1
B
1j
_
P
,
U
,x
1
x
1j
,
P
,
U
,x
1
x
1i
_
=
J
x
1i
+O
_
k

l=2
U
_
[x
1
x
l
[

__
.
Using Proposition (2.4), we can solve the above system to obtain the desired
estimates.
As a direct consequence of Proposition 2.4 and Lemma 2.5, we have
Proposition 2.6. Let (

(k, x), v
,k
(x)) be the map obtained in Proposi-
tion 2.3. Then for each xed i, we have
(2.47)
K(x)
x
il
=(p 1)
_

U
p2
,x
i
U
,x
i
x
il

,x
i
(p 1)

j=i
_

U
p2
,x
i
P
,
U
,x
j
U
,x
i
x
il
+O(
N1
e
(2+)d(x
i
,)/
)
+O
_

N1

j=i
U
1+
_
[x
i
x
j
[

_
+
N1

j=i
e
d(x
j
,)/
U
_
[x
i
x
j
[

__
.
By Proposition 2.6, Lemmas A.3 and A.4, we have
Proposition 2.7. Suppose that min
i=j
[x
i
x
j
[ > 2 max
1jk
d(x
j
, ).
Then
(2.48)
K(x)
x
ji
=
_

P
,
U
,x
j
n
U
,x
j
r
_
_
_
_
_

P
,
U
,x
j
n
U
,x
j
x
ji
_

P
,
U
,x
j
n
U
,x
j
r

ji
(x)
_
_
_
_
,
where r = [y x
l
[ and
ji
(x) satises
ji
(x) = O(e
d(x
j
,)/
).
3. Conley index and critical groups
Suppose that D
1
, . . . , D
k
are disjoint open sets compactly contained in
and satisfy min
i=j
d(D
i
, D
j
) > 2 max
1jk
max
x
j
D
j
d(x
j
, ). Let D = D
1

. . . D
k
. In order to prove that (1.1) has a k-peak solution with exactly one
peak in D
j
, j = 1, . . . , k, we need to prove that K(x) has a critical point x D.
A sucient condition to guarantee that K(x) has a critical point x D is that
the Conley index h(DK, D) is not trivial. Let
(3.1) X
j
(x
j
) =
_
_
_
_
_

P
,
U
,x
j
n
U
,x
j
x
j1
_

P
,
U
,x
j
n
U
,x
j
r
, . . . ,
_

P
,
U
,x
j
n
U
,x
j
x
jN
_

P
,
U
,x
j
n
U
,x
j
r
_
_
_
_
,
22 E. N. Dancer S. Yan
and X(x) = (X
1
(x
1
), . . . , X
k
(x
k
)). Since
P
,
U
,x
j
n
U
,x
j
r
_

P
,
U
,x
j
n
U
,x
j
r
is uniformly bounded in L
1
(), we may assume that there is a measure on
such that as 0 (at least for suitable subsequences),
P
,
U
,x
j
n
U
,x
j
r
_

P
,
U
,x
j
n
U
,x
j
r

j
.
It is easy to check that
_

d
j
= 1 and spt (
j
)

(x
j
) = y : y , [y
x
j
[ = d(x
j
, ). Thus, as 0, the limit of the vector eld in (3.1) is
_

x
j
y
[x y[
d
j
=
_

n(y) d
j
,
where n(y) is the outward unit normal of at y, since spt (
j
)

(x
j
) and
(y x
j
)/[y x
j
[ = n(y) for any y

(x
j
). From the result in [10], we see
_

((x
j
y)/[x
j
y[)d
j
d(x
j
, ), the Clarke gradient of d(x
j
, ).
Let o(1) denote any vector eld whose norm tends to zero as 0. Since
d(x, ) has no critical point on D
j
, it is easy to check that there is a c
0
> 0,
such that [X
j
(x
j
)[ c
0
for all x
j
D
j
. Using homotopy invariance, we deduce
from Proposition 2.7 that h(DK, D) = h(X +o(1), D) = h(X, D).
From the above discussion, we know that the limit of the vector eld in (3.1)
belongs to the Clarke gradient of the Lipschitz function d(x
j
, ). So in order
to see the eect of the domain geometry on the Conley index of the vector eld
in (3.1), it is much more convenient to discuss the Conley index of the Clarke
gradient of d(x
j
, ).
The aim of this section is to discuss the Conley index of the Clarke gradient of
any Lipschitz function. From now on, we assume that f(x) is a locally Lipschitz
function on R
m
.
We assume that T is a set of critical points of f satisfying that f is constant
on T. We also assume that T is isolated in the sense that there is an open
neighbourhood W of T such that 0 / f(x) for all x W T. It is convenient
to add a constant to f so f(x) = 0 for x T.
Since f(x) is set valued, we need to do more work before we can dene the
Conley index h(f, W). As in [8], for two open sets W
1
and W
2
satisfying
T W
2
W
2
W
1
W
1
W,
Multipeak Solutions 23
there is a locally Lipschitz map V (x) R
m
such that |V (x)| 1 for all x and
V (x), t(x))
1
, for all x W W
1
and t(x) f(x), (3.2)
V (x), t(x))
2
, for all x W W
2
and t(x) f(x), (3.3)
where
i
is a positive constant depending on W
i
. Suppose that x(t) is a solution
of dx(t)/dt = V (x(t)) satisfying x(t
1
) W and x(t
2
) W
1
for some t
1
and t
2
,
then
0 < [x(t
1
) x(t
2
)[ =

_
t
1
t
2
V (x(t)) dt

[t
1
t
2
[.
So if the ow x(t) satises x(t) W W
1
for all t (t
1
, t
2
), x(t
1
) W
and x(t
2
) W
1
for some t
1
and t
2
, then
f(x(t
2
)) f(x(t
1
)) =
_
t
2
t
1
df(x(t))
dt
dt c
0
(t
2
t
1
) c
0
= > 0,
since by [8], df(x(t))/dt c
0
> 0 for almost all t [t
1
, t
2
]. Now for xed W
1
,
we choose W
2
such that [f(y)[ /2 for all y W
2
. For this W, W
1
and W
2
,
we have a vector eld V (x) satisfying 3.2 and 3.3. Dene the Conley index
h(f, W) to be h(V (x), W).
Proposition 3.1. h(f, W) is well dened.
Proof. To prove that h(f, W) is well dened, we need to check that W is
an isolating neighbourhood of V and h(V (x), W) is independent of the choice
of V (x) satisfying (3.2) and (3.3).
First, we prove that W is an isolating neighbourhood of V . We argue by
contradiction. Suppose that there is a solution x(t) of dx(t)/dt = V (x(t)) such
that x(t) always stays in W and touch W at some time t

. If x(t) lies in WW
2
for all t t

, then by (3.3), we get

C f(x(t)) f(x(t

)) =
_
t
t

df(x(t))
dt
dt
1
(t

t) ,
as t . This is impossible. So we can nd a t
3
< t

such that x(t

3
) W
2
and thus f(x(t
3
)) /2. Choose t
1
(t
3
, t

) such that x(t

1
) W
1
and
x(t) stays in W W
1
for all t (t
1
, t

). From the above discussion, we have

f(x(t
1
)) f(x(t

)) . As a result,
f(x(t

)) f(x(t
1
)) f(x(t
3
)) = /2.
Similarly, we can nd t
4
> t
2
> t

, such that x(t

4
) W
2
, x(t
2
) W
1
and
f(x(t

)) f(x(t
2
)) . Thus
f(x(t
4
)) f(x(t
2
) f(x(t

)) 3/2.
This is a contradiction. So we have proved that W is an isolating neighbourhood
of V .
24 E. N. Dancer S. Yan
Since the class of V satisfying (3.2) and (3.3) is convex, the homotopy in-
variance of the usual Conley index shows that our denition is independent of
the choice of V .
Remark 3.2. By standard properties of the usual Conley index, h(f, W)
is independent of the choice of W and W
1
. Note that our construction seems to
bear some relation to Mischaikovs construction of the Conley index for multi-
valued maps.
Remark 3.3. If f(x) = f
1
(x
1
) +f
2
(x
2
), x
1
R
l
, x
2
R
ml
, then
h(f, W
1
W
2
) = h(f
1
, W
1
) h(f
2
, W
2
),
where W
1
R
l
and W
2
R
ml
. In fact, for each f
i
(x
i
), i = 1, 2, we choose
V
i
(x
i
) satisfying (3.2) and (3.3). By f(x) f
1
(x
1
) f
2
(x
2
) (in fact, equality
holds in this case), we see that V (x) = (V
1
(x
1
), V
2
(x
2
)) satises (3.2) and (3.3).
Thus, by the Conley index formula for products (as in [11]),
h(f, W
1
W
2
) = h(V
1
(x
1
) V
2
(x
2
), W
1
W
2
)
= h(V
1
(x
1
), W
1
) h(V
2
(x
2
), W
2
) = h(f
1
, W
1
) h(f
2
, W
2
).
Hence by the K unneth theorem for products in cohomology, the homotopy index
of f is nontrivial if the cohomology of f
1
and f
2
are nontrivial (for the same eld
of coecients). This is where the Conley index method has advantages.
Definition 3.4. Let f satisfy the conditions mentioned in the begining of
this section. We dene the critical group C(f, T) of f on T to be the cohomology
of h(V, W) for V and W satisfying (3.2) and (3.3). As usual, we choose a eld
as the coecients.
Now we want to prove an alternative formula for the critical groups, which
in practice is easier to calculate with. Let | be a neighbourhood of T.
Proposition 3.5. We have C(f, T) = H

(f
0
|, (f
0
T) |), where
f
c
= x : f(x) c.
Proof. To prove our claim above, we make a special choice of W. We follow
the idea in [13]. By using an inductive procedure in the Changs construction,
we can construct V (x) locally Lipschitz on | T, so that |V (x)| 1 and
V (x), t(x)) > 0 on | T and V (x), t(x)) > 0 on any compact subset S of
|T, for any t(x) f(x). Here depends on S. Choose a small neighbourhood
Z of T and suppose that > 0 is relatively small.
Let x(t) be a solution of x

(t) = V (x(t)), x(0) ZT. If f(x(0)) 0, then

since f(x(t)) is strictly decreasing in t, we see there is a positive lower bound for
the distance between T and x(t) : t 0. As a result, df(x(t))/dt < 0.
So we see that x(t) hits f
1
(). Backwards in time, if there is a positive lower
Multipeak Solutions 25
bound for the distance between T and x(t) : t 0, then x(t) hits f
1
(). So
we have proved that either x(t) hits both f
1
() and f
1
(), or the ow goes
to T. A similar result is also true for the case f(x(0)) 0. Now we form a set W
by taking all these ow lines starting in Z and choosing the part of the ow line
till it hits x : f(x) = or hits T, together with T. It is easily seen that W
is a neighbourhood of T and contains no other critical points of f. Moreover,
(W, W f
1
()) is an index pair in the sense of Conley.
Now we calculate h(V (x), W). By our construction, we see that the exit
set for the ow of V on W is W f
1
(). Thus,
h(V (x), W) = [W/(W f
1
())].
From the above discussion, we know that W = f

W can be deformed into

f
s
W along the ow line for V for any s (0, ). Hence, noting that f
s
W
decreases to f
0
W, we see by results on direct limits for Alexander cohomology
(see p. 238 in [25]) that
H([W/(W f
1
())]) = H(W, W f
1
()) = H(f
0
W, W f
1
())
= H
_
f
0
W, f
0
W T
_
,
since f
0
W T can be deformed into f
1
() W. Thus the result follows.
By the homotopy invariance of the degree, we can also use the vector eld
V (x) to dene the degree of the Clarke gradient of a Lipschitz function.
Definition 3.6. Let f satisfy the conditions mentioned in the begining
of this section. We dene deg(f, W, 0) = deg(V, W, 0), where V and W sat-
isfy (3.2) and (3.3).
For a smooth function, we have
deg(Df, W, 0) =
m

i=0
(1)
i
rank C
i
(f, x
0
),
where x
0
is an isolated critical point f and W is a small neighbourhood of x
0
.
Our next result shows that the above relation is still true for a Lipschitz function.
Proposition 3.7. Let f be a Lipschitz function and let x
0
be the unique
critical point of f in a neighbourhood W of x
0
. Then
deg(f, W, 0) =
m

i=0
(1)
i
rank C
i
(f, x
0
).
Proof. First, by Rademachers Theorem, f is dierentiable almost every-
where. Besides, by Proposition 2.2.2 in [10], if f is dierentiable at x
0
, then
Df(x
0
) f(x
0
).
26 E. N. Dancer S. Yan
Let C

0
(B
1
(0)) be a function with 0 1 and
_
R
N
(y) dy = 1. Let

(x) =
N
(
1
y) for > 0 small. Dene
f

(x) =
_
R
N

(y)f(x y) dy.
Then by the dominated convergence theorem, we have
Df

(x) =
_
R
N

(y)D
x
f(x y) dy.
Let V and W satisfy (3.2) and (3.3). Since Df(x) f(x) almost everywhere
and V (x) is continuous, we have
D
x
f(x y), V (x)) = D
x
f(x y), V (x y)) +o(1) /2
for almost y B

(x) if > 0 is small. As a result,

Df

(x), V (x)) =
_
R
N

(y)D
x
f(x y), V (x)) dy
=
_
B

(x)

(y)D
x
f(x y), V (x)) dy

2
> 0,
So we obtain
deg(f, W, 0) = deg(V, W, 0) = deg(Df

, W, 0)
and
h(f, W) = h(V, W) = h(Df

, W).
Now since f

is a smooth function, we can argue in exactly the same way as

in [13, p. 14] to nd that
deg(Df

, W, 0) =
m

i=1
(1)
i
rank H
i
(h(Df

, W)).
Thus the result follows.
Remark 3.8. In [23], Li and Nirenberg used classical degree theory to study
the single peak solution. The main result in [23] states that if A is an open set
compactly contained in and d(x, ) is dierentiable on A, then (1.1) has
a single peak solution with its peak inside A provided deg(Dd(x, ), A, 0) ,= 0.
It is worth pointing out that for an isolated critical point x
0
of d(x, ) , it is not
always possible to nd a small neighbourhood A of x
0
such that x
0
is the unique
critical point of d(x, ) in A and d(x, ) is dierentiable on A. For example,
if is a symmetric dumbell, then d(x, ) is not dierentiable on the segment
joining the centers of the balls. Thus any A containing one of the center of the
ball (the local maximum point of d(x, )), or the middle of the handle (the
saddle point of d(x, )) must contains the whole segment joining the centers
of the balls if d(x, ) is dierentiable on A. Here we generalize the classical
Multipeak Solutions 27
degree theory to the Clarke gradient of any Lipschitz function f and the above
proposition shows that if deg(f, W, 0) ,= 0, then C(f, x
0
) is nontrivial.
Now we give some examples where the critical points have nontrivial critical
groups.
Proposition 3.9. Suppose that T is a set of local maximum points of f such
that f is constant on T and T is an isolated set of critical points of f. In this
case, C
i
(f, T) = H
Ni
(T), where the homology is non-reduced and is Steenrod
homology as in [25]. In particular, C
N
(f, T) is nontrivial.
Proof. This follows from Theorem 11.15 in [25] and Proposition 3.5 here.
Proposition 3.10. Suppose that x
0
= 0 is a saddle point in the following
sense: there is an integer l 1 and an l-dimensional C
1
manifold Y and an
m l dimensional C
1
manifold Z such that Y Z = 0, R
m
= T
0,Y
T
0,Z
,
where T
0,Y
and T
0,Z
are the tangent spaces of Y at 0 and Z at 0 respectively,
f(x) > f(0) if x B

(0) Z 0, f(x) < f(0) if x B

(0) Y 0 for
some small > 0. Then if x
0
= 0 is the unique critical point of f in B

(0),
C
l
(f, x
0
) ,= 0.
Proof. It is easy to see that there is a C
1
local dieomorphism : R
m

R
m
such that (Y ) and (Z) are subspaces near the origin. Then the proof of
the claim is similar to that of Theorem 2.1 in [24].
We can relax the above denition for saddle point a little bit so that it is
much easier to check in practice.
Proposition 3.11. Suppose that x
0
= 0 is a saddle point in the following
sense: there is an integer l 1 and an l-dimensional C
1
manifold Y and an
m l dimensional C
1
manifold Z such that Y Z = 0, R
m
= T
0,Y
T
0,Z
,
where T
0,Y
and T
0,Z
are the tangent spaces of Y at 0 and Z at 0 respectively,
f(x) f(0) if x B

(0)Z, f(x) f(0) if x B

(0)Y for some small > 0.

Then if x
0
= 0 is the unique critical point of f in B

(0), we have C
l
(f, x
0
) ,= 0.
Proof. We just need to nd an l-dimensional manifold Y

and an m l
dimensional manifold Z

such that Y

= 0, R
m
= T
0,Y
T
0,Z
, where
T
0,Y
and T
0,Z
are the tangent spaces of Y

at 0 and Z

at 0 respectively,
f(x) > f(0) if x B

(0) Z

0, f(x) < f(0) if x B

(0) Y

0 for some
small > 0. This claim follows if we can choose a C
1
vector eld V (x) satisfying
V (0) = 0, DV (0) = 0, V (x), t(x)) > 0 for x B

(0) 0 and t(x) f(x). In

fact, let Y

= x = x(, x
0
), x
0
Y , where x(t, x
0
) is a solution of
x

(t) = V (x(t)), x(0) = x

0
28 E. N. Dancer S. Yan
and > 0 is a small constant, and let Z

= x = x(, x
0
), x
0
Z, where x(t, x
0
)
is a solution of
x

(t) = V (x(t)), x(0) = x

0
and > 0 is a small constant. Then by using DV (0) = 0, it is easy to check
that T
0,Z
= T
0,Y
and T
0,Z
= T
0,Z
. Thus Y

and Z

satisfy the requirements.

To prove the existence of such vector eld V (x), since 0 is an isolated crit-
ical point of f, by using the construction in [8], we know that there is vec-
tor eld V

(x) which is locally C

1
in B

(0) 0 and satisfying |V

(x)| 1,
V

(x), t(x)) > 0 for x B

(0) 0 and t(x) f(x). For each integer

j > 0, we can nd a increasing sequence of L
j
> j such that [DV

(x)[ L
j
for x B

(0) B
1/j
(0). Let (t) is a C
1
function satisfying (0) = 0, (t) > 0
for t > 0 and

(t) 1/L
2
j+1
for t j
1
. Let V (x) = ([x[)V

(x). Then for

any x B

(0) 0, we can nd a j, such that x B

1/j
(0) B
1/(j+1)
(0). Thus,
[DV

(x)[ L
j+1
. As a result, we have
[DV (x)[ ([x[)[DV

(x)[ +[D([x[)[
[x[
L
2
j+1
L
j+1
+ 1/L
2
j+1
= O([x[).
So V (x) is the vector eld we need.
Remark 3.12. According to the above results, it is easy to check that the
critical groups of the critical points in examples 1.12, 1.13 and 1.14 given in [21]
are nontrivial.
Remark 3.13. In [33], a point x
0
is called a nondegenerate peak point if
_

e
zx
0
,a
(z x
0
) d
x
0
= 0
and the matrix
__

e
zx
0
,a
(z
i
x
0,i
)(z
j
x
0,j
) d
x
0
_
is nonsingular, where a is some point in R
N
,
x
0
is a weak limit of
e
|zx
0
|/
__

e
|zx
0
|/
_
1
as 0. It is proved in [28] that a point x
0
is a nondegenerate peak point
if and only if x
0
int(co

(x
0
)). Here, we want to point out a nondegenerate
peak point is a strictly local maximum point of the distance function. To see
this, let e be any unit vector in R
N
. We claim that there is a > 0, independent
of e, such that y, e) for some y

(x
0
). We argue by contradiction.
Suppose that there is a sequence of unit vector e
i
such that y, e
i
) o(1) for any
y

(x
0
) as i . Assume e
i
e
0
. Hence, y, e
0
) 0 for any y

(x
0
).
By translation and rotation, we may assume that x
0
= 0 and e
0
= (0, . . . , 0, 1).
Since 0 int(co

(0)), we can nd a point y

(0) satisfying y, e
0
) > 0.
Multipeak Solutions 29
Otherwise, co

(0) y
N
0, which contradicts 0 int(co

(0)). Thus
our claim follows. So, for t > 0 small (independent of e), we have
[y te[
2
= [y[
2
2ty, e) +t
2
< [y[
2
,
which implies d(te, ) [y te[ < [y[ = d(0, ).
Before we close this section, let us look at the following examples.
Example 3.14. Now we give an example where the critical point of the dis-
tance function is not the kind of saddle point dened in [9], but the corresponding
critical group is nontrivial.
Let g(x
1
, x
2
) = 1 + r
2
cos a, where > 0 is a small constant such that
normal line of the surface x
3
= g(x
1
, x
2
) at (x
1
, x
2
, g(x
1
, x
2
)) does not intersect
with the normal line of this surface at other point within the region [x
3
[ 2,
a is prime, r = (x
2
1
+ x
2
2
)
1/2
and is the polar angle of x
1
x
2
plane. Dene
a domain in R
3
as follows:
= (x
1
, x
2
, x
3
) : [x
3
[ g(x
1
, x
2
), r R,
where R > 0 is a large constant. Then it is easy to check is invariant under
rotation of 2/a in x
1
x
2
plane.
Firstly we claim that x = 0 is an isolated critical point of d(x, ). In fact, let
x ,= 0 be a point in a small neighbourhood of the origin. If x is not in the x
1
x
2
plane, then

(x) contains just one point and thus is not a critical point. If x
is in the x
1
x
2
plane, then

(x) contains exactly two points y = (y

1
, y
2
, y
3
)
and y

= (y
1
, y
2
, y
3
) with y
3
= g(y
1
, y
2
) and y
2
1
+ y
2
2
,= 0. Direct calculation
shows that Dg(y
1
, y
2
) ,= 0 if y
2
1
+ y
2
2
,= 0. But y x = c
0
(Dg(y
1
, y
2
), 1) and
y

x = c
0
(Dg(y
1
, y
2
), 1). As a result, y x ,= (y

x) which implies
0 / co(

(x) x).
Secondly, we show x = 0 is not the kind of saddle point dened in [9].
Suppose that X
1
and X
2
are two subspaces in R
3
such that R
3
= X
1
X
2
,
max
xX
1
B

(0)
d(x, ) < min
xX
2
B

(0)
d(x, ); max
xX
1
B

(0)
d(x, ) <
min
xX
2
B

(0)
d(x, ). For any x with polar angle satisfying cos a 0
and r ,= 0, it is easy to check d(x, ) < g(x
1
, x
2
) 1. On the other hand, for
any small > 0, there exists x X
2
with its polar angle satisfying cos a 0
and [x[ = . So we see d(x, ) < 1. As a result, max
xX
1
B

(0)
d(x, )
d(0, ) > d(x, ) min
xX
2
B

(0)
d(x, ). This is a contradiction.
In addition, let x = (x
1
, x
2
, 0) be a point such that cos a > c
0
> 0. It is easy
to check d(x, ) > 1. This shows that x = 0 is not a local maximum point.
Finally, to calculate the Conley index, we see easily that we can choose the
approximating vector elds to preserve the symmetry of rotation of 2/a in the
x
1
x
2
plane. Hence we can apply the results in Remark (ii) on page 672 of [12]
(using a remark on page 14 in [13] to remove a side condition). We see that
30 E. N. Dancer S. Yan
if the Clarke gradient of the distance function had Conley index trivial at 0,
then it would also have to have trivial homology with Z
a
coecients on the one
dimensional subspace x
1
= x
2
= 0, which is the xed point set of the rotation.
However, this is impossible since 0 is a maximum point of the distance function
on this space.
Using a similar argument, we can also prove that in this example, d(x, )
does not satisfy the conditions in Proposition 3.11.
Example 3.15. Let
1
be a bounded domain in x
1
x
3
plane such that
x
1
> 0 if (x
1
, x
3
)
1
. Suppose that x

= (x

1
, x

3
)
1
is a saddle point of the
distance function in the strong sense, that is, for xed x
1
, d(x,
1
) is increasing
in x
3
> x

3
and decreasing in x
3
< x

3
, while for each xed x
3
, d(x,
1
) is
increasing in x
1
< x

1
and decreasing in x
1
> x

1
. The three dimensional domain
is obtained by rotating
1
around the x
3
-axis. So has an isolated critical
set T = (x

1
cos , x

1
sin , x

3
), [0, 2].
Let d
0
= d(x

,
1
), x

= (x

1
, x

3
). It is easy to see the pair (d
d
0

(x

), (d
d
0
x

)B

(x

)) can be deformed into (I, I), where I = (x

1
, x

3
) :
x
1
[x

1
, x

1
+]. As a result, we have
C(d, T) = H

(S I, S I) = H

(S) H

(I, I),
where S is the unit circle in R
2
. Hence, C(d, T) is nontrivial.
To calculate the degree of Dd, we use the symmetries again. By using a S
1
equivariant approximation and a theorem of Nussbaum [29], we see that the
degree of Dd is the same as on the symmetric subspace x
1
= x
2
= 0. But the
intersection of this subspace with is empty and hence the degree is zero.
We can also calculate the degree of Dd by proving the natural analogue of
Proposition 3.7.
Remark 3.16. We can obtain even more complicated four dimensional ex-
amples by rotating the sets in Example 3.14 around an axis (much as in Exam-
ple 3.15).
4. Proof of the main results
In this section, we will use the results in Section 2 to get some existence
result for (1.1).
Proof of Theorem 1.1. Let X
j
(x
j
) be the vector eld dened in Section 3.
Since it is not clear whether X
j
(x
j
) converges uniformly in a neighbourhood of
D
j
, we cannot conclude immediately that h(X(x
j
), D
j
) = h(d(x, ), D
j
).
To check h(X(x
j
), D
j
) = h(d(x, ), D
j
), we choose the vector eld V (x
j
)
satisfying (3.2) and (3.3). We claim that V (x
j
), X
j
(x
j
)) > 0 in a neigh-
bourhood of D
j
. By the homotopy invariance of the Conley index, we then
Multipeak Solutions 31
obtain h(X(x
j
), D
j
) = h(V (x
j
), D
j
). Thus
H

(h(X(x
j
), D
j
)) = H

h(V (x
j
), D
j
)) = C(d
c
j
D
j
, (d
c
j
D
j
) T
j
).
By assumption, H

(h(X(x
j
), D
j
)) is nontrivial. But
h(X, D) = h(X(x
1
), D
1
) . . . h(X(x
k
), D
k
).
Thus it follows from the K unneth formula that H

(h(X, D)) is nontrivial. So

h(X, D) is nontrivial and the result follows.
It remains to prove the claim. We argue by contradiction. Suppose that there
is a sequence x
j,m
in a neighbourhood of D
j
, such that V (x
j,m
), X
j
(x
j,m
))
c
0
0 as m . Assume x
j,m
x
0
. Since

m
(y) =:
P
,
U
,x
j,m
n
U
,x
j,m
r
_

P
,
U
,x
j,m
n
U
,x
j,m
r
is uniformly bounded in L
1
(), we may assume that there is a measure on
such that as m ,
m
. Thus,
V (x
j,m
), X
j
(x
j,m
)) =V (x
0
), X
j
(x
j,m
)) +o(1)
=
_

_
y x
0
[y x
0
[
, V (x
0
)
_

m
(y) dy
+
_

_
y x
j,m
[y x
j,m
[

y x
0
[y x
0
[
, V (x
0
)
_

m
(y) dy +o(1)

__

y x
0
[y x
0
[
d, V (x
0
)
_
= c
0
0.
It is easy to check that
_

d = 1 and spt()

(x
0
). Thus,
_

(y
x
0
)/[y x
0
[ d d(x
0
, ). This is a contradiction to (3.2).
Proof of Theorem 1.2. First we claim that in a strictly convex domain,
d(x, ) has exactly one critical point. Thus this critical point is the global
maximum point of d(x, ). We argue by contradiction. Suppose that d(x, )
has two critical points x
1
and x
2
. By translation and rotation, we may assume
that x
1
= 0, x
2
= (l, 0, . . . , 0) and d
1
= d(x
1
, ) d
2
= d(x
2
, ). From [10],
we know that
d(x, ) co
_
y x
[y x[
: y

(x)
_
,
where co denotes the convex hull. Since 0 d(x
1
, ), we see that

(x
1
) y = (y
1
, . . . , y
N
) R
N
, y
1
0 , = .
Take y

(x
1
) y = (y
1
, . . . , y
N
) R
N
, y
1
0. It is easy to check that
T
y
( y) ,= . This is a contradiction.
32 E. N. Dancer S. Yan
Since is convex, using the moving plane method of Gidas, Ni and Nirenberg,
we know that there is a d
0
> 0, such that any local maximum point x
0
of the
solution of (1.1) satises d(x
0
, ) d
0
. Suppose that (1.1) has a solution of
the form (1.4) satisfying (1.5). Since there is always a local maximum point in
a small neighbourhood of x
,j
, we know that d(x
,j
, ) d
0
. Arguing as in [17],
we obtain min
i=j
[x
,i
x
,j
[ d
0
.
Without loss of generality, we assume d(x
,1
, ) = min
1jk
d(x
,j
, ).
We also assume x
,j
x
j
as 0. As in Section 2, we have
K(x

)
x
ji
= 0.
Since is convex, it follows from Proposition 2.6 and Lemma A.7 that the above
relation is equivalent to
0 =(p 1)
_

U
p2
,x
,1
U
,x
,1
x
1l

,x
,1
(4.1)
(p 1)
k

j=2
_

U
p2
,x
,1
U
,x
,j
U
,x
,1
x
1l
+O(
N1
e
(2+)d(x
,1
,)/
)
+o
_

N1
k

j=2
U
_
[x
,1
x
,j
[

__
.
Thus, by Lemma A.4, we have
(4.2) 0 =
_

P
,
U
,x
1
n
U
,x
1
r
__

y x
1
[y x
1
[
d +o(1)
_
+c
0

N1
k

j=2
U
_
[x
,1
x
,j
[

__
x
1
x
j
[x
1
x
j
[
+o(1)
_
+O(
N1
e
(2+)d(x
,1
,)/
) +o
_

N1
k

j=2
U
_
[x
,1
x
,j
[

__
,
where is a measure satisfying
_

d = 1 and spt

(x
1
).
Since d(x
1
, ) = min
1jk
d(x
j
, ) and min
2jk
[x
j
x
1
[ > 0, we know
x
1
is not the global maximum point of d(x, ) and thus is not a critical point
of d(x, ). So

y x
1
[y x
1
[
d

> 0.
By (4.2), we obtain
Multipeak Solutions 33
(4.3) 0 =
_

P
,
U
,x
1
n
U
,x
1
r
_

y x
1
[y x
1
[
d

2
+o(1)
_
+c
0

N1
k

j=2
U
_
[x
,1
x
,j
[

___
x
1
x
j
[x
1
x
j
[
,
_

y x
1
[y x
1
[
d
_
+o(1)
_
+O(
N1
e
(2+)d(x
,1
,)/
) +o
_

N1
k

j=2
U
_
[x
,1
x
,j
[

__
,
We claim that
(4.4)
_
x
1
x
j
[x
1
x
j
[
,
_

y x
1
[y x
1
[
d
_
> 0.
So by Lemma A.3, we see (4.3) is impossible and thus the result follows.
It remains to check (4.4). Without loss of generality, we assume that x
1
= 0
and x
2
= (l, 0, . . . , 0). Arguing as in the proof of the uniqueness of critical point
for d(x, ), we obtain

(x
1
) y = (y
1
, . . . , y
N
) R
N
, y
1
< 0,
which clearly implies (4.4).
Finally, we give an example which shows that if the critical group of an
isolated critical point x
0
of d(x, ) is trivial, there may be no positive solution
of (1.1) such that one of its local maximum point is close to x
0
.
Example 4.1. Let =
1

2
, where
1
= B
1
(0) x = (x
1
, . . . , x
N
)
R
N
, x
1
0,
2
is any open domain such that is smooth and the reection
of x
1
t about x
1
= t is contained in for all t , where > 0 is
a constant. Then x
0
= 0 is a critical point of d(x, ) and the corresponding
critical group is trivial. By using the moving plane method of Gidas, Ni and
Nirenberg, we conclude that any positive solution of (1.1) is increasing in the
direction x
1
till x
1
= and thus there is no local maximum point in x
1
.
5. Remark on the Neumann problem
Let be a bounded domain in R
N
with smooth boundary. Consider the
following Neumann problem:
(5.1)
_

2
u +u = u
p1
in ,
u > 0 in ,
u
n
= 0 in ,
where is a small positive number, n is the unit outward normal of at y,
2 < p < 2N/(N 2) if N 3 and 2 < p < if N = 2.
34 E. N. Dancer S. Yan
We denote P
,,N
v the solution of the following problem:
_

2
u +u = [v[
p2
v y ,
u
n
= 0 y ,
u H
1
().
By the maximum principle, we know P
,,N
U
,z
> 0.
For any x
i
, i = 1, 2, . . . , k, dene
E

,x,k
=
_
v H
1
() : P
,,N
U
,x
i
, v)

= 0,
_
P
,,N
U
,x
i

i
, v
_

= 0, i = 1, 2, . . . , k
_
,
where
i
is any unit vector in R
N
if x
i
;
i
is any tangent vector of at x
i
.
Let H(x) be the mean curvature function of . We use I or J to denote
a nite index set and use [I[ to denote the number of points I contains. Using
the techniques in Section 2 and the results in Section 3, we can get the following
result:
Theorem 5.1. Let I and J be two nite index sets where one of the sets may
be empty. Suppose that x
i
, i I, are dierent critical points of H(x) with
C(H, x
i
) nontrivial and z
j
, j J, are dierent critical points of d(x, )
with C(d, z
j
) nontrivial. If
min
iI,l,jJ,l=j
([x
i
z
j
[, [z
l
z
j
[) > 2 max
jJ
d(x
j
, ),
then there exists an
0
> 0, such that for each (0,
0
], (5.1) has at least one
solution of the form
(5.2) u

iI

i
P
,,N
U
,x
i
+

jJ

j
P
,,N
U
,z
j
+v

where, as 0,
i
1, x
i
x
i
, x
i
, i I,
j
1, z
j
z
j
, j J,
v

,x

,|I|+|J|
and |v|
2

= o(
N
).
Appendix A. Basic estimates
In this section, we present some basic estimates needed in the proof of the
main results. First, let us recall the well known fact on the asymptotic behaviours
of U(y):
lim
|y|
[y[
(N1)/2
e
|y|
U(y) = c
0
> 0.
From now on, we always assume that d(x, )/ M for some large constant
M > 0. Let
,x
= U
,x
P
,
U
,x
. Then
,x
satises
(A.1)
_

2

,x
+
,x
= 0 y ,

,x
= U
,x
y .
Multipeak Solutions 35
We denote

,x
=
_

U
p1
,x

,x
.
We have the following estimate for
,x
.
Lemma A.1. For any > 0, there exist C
2
> C
1
> 0, such that
(A.2) C
1

N
e
(2+)d(x,)/

,x
C
2

N
e
(2)d(x,)/
.
Lemma A.2. There is a > 0, such that
_

2
,x
U
p2
,x
= O(e
d(x,)/

,x
), (A.3)
_

p
,x
= O
_
e
d(x,)/

,x
_
. (A.4)
Dene

,x
=
2
_

P
,
U
,x
n
U
,x
r
,
where r = [y x[ and n is the outward unit normal to at y.
Lemma A.3. For any > 0, there exist c
1
> c
0
> 0, such that
c
0

N1
e
(2+)d(x,)/

,x
c
1

N1
e
(2)d(x,)/
.
Lemma A.4. We have
_

U
p2
,x
U
,x
x
i

,x
=
2
_

P
,
U
,x
n
U
,x
x
i
+O(
N1
e
pd(x,)/
).
Moreover, if B
d(x,)
(x) contain exactly one point q, then
N

i=1
_

U
p2
,x
U
,x
x
i

,x

i
c
0

N1
e
(2+)d(x,)/
,
for any > 0, where is the outward unit normal to at q and c
0
> 0 is
a constant.
Lemma A.5. For i ,= j, we have
(A.5)
_

U
p1
,x
i

,x
j
= O
_

N
N

j=1
e
(2+)d(x
j
,)/
+
N

i=j
e
(1+)|x
i
x
j
|/
_
.
The proof of the above lemmas can be found in [16]. So we omit the proof
of these lemmas.
36 E. N. Dancer S. Yan
Lemma A.6. There is a > 0, such that for i ,= j,
_

U
,x
i
U
p2
,x
j

,x
j
= O
_

N
e
d(x
j
,)/

i=j
e
|x
i
x
j
|/
_
.
Proof. Choose > 0 small enough so that p2 > 0. Since
,x
j
U
,x
j
,
we have
_

U
,x
i
U
p2
,x
j

,x
j
C
_

e
|yx
i
|/
U
p2
,x
j

,x
j
Ce
|x
j
x
i
|/
_

e
|yx
j
|/
U
p2
,x
j

,x
j
Ce
|x
j
x
i
|/
_

U
p2
,x
j

,x
j
Ce
|x
j
x
i
|/
e
d(x
j
,)/
_

U
p2
,x
j
= O
_

N
e
d(x
j
,)/

i=j
e
|x
i
x
j
|/
_
.
Lemma A.7. Suppose that is convex. For i ,= j, if d(x
l
, ) > 0,
l = i, j, we have
(A.6)
_

U
p1
,x
i

,x
j
= O(
N
e
(1+)|x
i
x
j
|/
).
Proof. Let G(z, y) be the Greens function of
2
u + u on subject to
Dirichlet boundary conditions. Then

,x
j
(y) =
2
_

G(z, y)
n
U
,x
j
(z) dz,
where n is the outward unit normal of at y. But for any > 0 small, we have

G(z, y)
n

Ce
(1)|zy|/
.
Since is convex, we have that there is a > 0 such that [y z[ + [z x[
(1 + )[x y[ for any z and x, y with d(x, ) and d(y, ) .
As a result,

,x
j
(y)
_

e
(1)|zy|/
e
|zx|/
dz Ce
(1+)|x
j
y|/
,
and the result follows.
Multipeak Solutions 37
References
[1] A. Bahri, Critical points at innity in some variational problems, Res. Notes Math.
182 (1989), Longman-Pitman.
[2] V. Benci and G. Cerami, The eect of the domain topology on the number of positive
solutions of nonlinear problems, Arch. Rational Mech. Anal. 114 (1991), 7993.
[3] V. Benci, G. Cerami and D. Passaseo, On the number of positive solutions of some
nonlinear elliptic problems, Nonlinear Analysis A Tribute in Honour of G. Prodi,
Quaderni Scuola Norm. Sup. Pisa (A. Ambrosetti and A. Marino, eds.), Scu. Norm. Sup.
Pisa, Pisa, 1993, pp. 93107.
[4] G. Cerami and D. Passaseo, Multiplicity results of positive solution for nonlinear
elliptic problems in exterior domain with rich topology, Nonlinear Anal. 18 (1992),
109119.
[5] , Existence and multiplicity results for semilinear elliptic Dirichlet problems in
exterior domains, Nonlinear Anal. 24 (1995), 15331547.
[6] D. Cao, E. N. Dancer, E. Noussair and S. Yan, On the existence and prole
of multi-peaked solutions to singularly perturbed semilinear Dirichlet problems, Discrete
Contin. Dynam. Systems 2 (1996), 221236.
[7] D. Cao and E. Noussair, Multi-peak solutions for a singularly perturbed semilinear
elliptic problem, preprint.
[8] K. C. Chang, Variational methods for non-dierentiable functionals and their applica-
tions to partial dierential equations, J. Math. Anal. Appl. 80 (1981), 102129.
[9] , Innite Dimensional Morese Theorey and Multiple Solutions Problems, Birk-
hauser, Boston, 1993.
[10] F. H. Clarke, Optimization and Nonsmooth Analysis, John Wiley and Sons, 1983.
[11] C. Conley, Isolated invariant sets and the Morse index, CBMS Regional Conference
No. 38, Amer. Math. Soc., Providence, 1978.
[12] E. N. Dancer, Symmetries, degree, homotopy indices and asymptotically homogeneous
problems, Nonlinear Anal. 6 (1982), 667686.
[13] , Degenerate critical points, homotopy indices and Morse inequalities, J. Reine
Angew. Math. 350 (1984), 122.
[14] E. N. Dancer and J. Wei, On the eect of domain topology in a singular perturbation
problem, Topol. Methods Nonlinear Anal. 11 (1998), 227248.
[15] E. N. Dancer and S. Yan, Multipeak solutions for a singularly perturbed Neumann
problem, Pacic J. Math. 189 (1999), 241262.
[16] , Singularly perturbed elliptic problem in exterior domains, J. Dierential Int.
Equations (to appear).
[17] , A singularly perturbed elliptic problem in bounded domains with nontrivial
topology, Adv. Dierential Equations 4 (1999), 347368.
[18] M. delPino, P. Felmer and J. Wei, Multiple peak solutions of singular perturbation
problems; preprint.
[19] B. Gidas, W. M. Ni and L. Nirenberg, Symmetry and related properties via the
maximum principle, Comm. Math. Phys. 68 (1979), 209243.
[20] D. Gilbarg and N. S. Trudinger, Elliptic Partial Dierential Equations of Second
Order, Springer-Verlag, 1983.
[21] M. Grossi and A. Pistoia, On the eect of critical points of distance function in
superlinear elliptic problems; preprint.
[22] M. K. Kwong, Uniqueness of positive solutions of u+u = u
p
in R
n
, Arch. Rational
Mech. Anal. 105 (1989), 243266.
38 E. N. Dancer S. Yan
[23] Y. Y. Li and L. Nirenberg, The Dirichlet problem for singularly perturbed elliptic
equation, Comm. Pure Appl. Math 51 (1998), 14451490.
[24] J. Liu, The Morse index of a saddle point, Systems Sci. Math. Sci. 2 (1989), 3239.
[25] W. Massey, Homology and Cohomology Theory, Marcel Dekker, Basel, 1978.
[26] R. Molle and D. Passaseo, On the behaviour of the solutions for a class of nonlinear
elliptic problems in exterior domains, Discrete Contin. Dynam. Systems (to appear).
[27] W. M. Ni and J. Wei, On the location and prole of spike-layer solutions to singularly
perturbed semilinear Dirichlet problems, Comm. Pure Appl. Math. 48 (1995), 731768.
[28] E .S. Noussair and S. Yan, The eect of the domain geometry in singular perturbation
problems, Proc. London Math. Soc. 76 (1998), 427452.
[29] R. Nussbaum, Some generalizations of the Borsuk-Ulam theorem, Proc. London Math.
Soc. 35 (1977), 136158.
[30] O. Rey, The role of the Greens function in a non-linear elliptic equation involving the
critical Sobolev exponent, J. Funct. Anal. 89 (1990), 152.
[31] J. Wei, On the construction of single-peaked solutions to a singularly perturbed semi-
linear Dirichlet problem, J. Dierential Equations 129 (1996), 315333.
[32] , On the construction of single-peaked solutions to a singularly perturbed semi-
linear Dirichlet problem, J. Dierential Equations 129 (1996), 315333.
[33] , Necessary geometric conditions for two-peaked solutions, Manuscripta Math.
96 (1998), 113131.
[34] , On the eect of the domain geometry in singular perturbation problems, pre-
print.
Manuscript received July 10, 1999
E. Norman Dancer and Shusen Yan
School of Mathematics and Statistics,
University of Sydney,
NSW 2006, AUSTRALIA
E-mail address: normd@maths.usyd.edu.au, shusen@maths.usyd.edu.au
TMNA : Volume 14 1999 N
o
1