STABILITY AND SYMMETRY OF SOLITARY-WAVE SOLUTIONS

TO SYSTEMS MODELING INTERACTIONS OF LONG WAVES

John Albert and Felipe Linares
Abstract. We consider systems of equations which arise in modelling strong interactions
of weakly nonlinear long waves in dispersive media. For a certain class of such systems, we
prove the existence and stability of localized solutions representing coupled solitary waves
travelling at a common speed. Our results apply in particular to the systems derived by
Gear and Grimshaw and by Liu, Kubota, and Ko as models for interacting gravity waves in
a density-stratied uid. For the latter system, we also prove that any coupled solitary-wave
solution must have components which are all symmetric about a common vertical axis.
1. Introduction.
Model equations for long, weakly nonlinear waves in uids are typically derived by
expanding the full equations of motion to rst order in a small parameter determining
the size of the wave amplitude and inverse wavelength. (The use of one small parameter
to describe these two small quantities implicitly assumes a certain balance between them.)
The solutions of the model equations describe the slow evolution, due to weak dispersive
and nonlinear eects, of a wave which in the linear, non-dispersive limit corresponds to a
mode of a linear eigenvalue problem.
The well-known Korteweg-de Vries equation, for example, was derived in this way by
Benney [7] as a model for internal waves in a vertically stratied uid. To zeroth order in ,
the full equations of motion are separable in the horizontal and vertical space coordinates,
giving rise to a Sturm-Liouville problem in the vertical coordinate and the linear wave
equation u
tt
c
2
j
u
xx
= 0 in the horizontal coordinate, where the wavespeed c
j
corresponds
to the jth eigenvalue of the Sturm-Liouville problem. Each individual eigensolution of
the Sturm-Liouville problem thus gives rise to a wave with xed vertical structure and
horizontal speed c
j
. The Korteweg-de Vries equation describes the eects of weak non-
linearity and weak dispersion on such a wave in the case when the horizontal motion is
unidirectional.
2
In this paper, we consider systems of equations which have been derived as models
for the interaction of two (or more) long waves, each of which corresponds to a dierent
underlying mode or vertical structure. Such systems generally take the form
h
t
+D
1
(N(h) Lh)
x
= 0, (1.1)
where h is an R
n
-valued function of x and t, D is a positive n n diagonal matrix, N is
the gradient of a homogeneous function N : R
n
R, and L is a Fourier multiplier operator
which acts self-adjointly on the Sobolev space in which (1.1) is posed. The interesting case
is when the wavespeeds corresponding to two dierent modes have nearly the same value,
so that the modes interact on a time scale long enough for nonlinearity and dispersion to
have a signicant eect.
In particular we are interested in solitary waves, or localized travelling-wave solutions
of (1.1) of permanent form. More precisely, by a solitary wave we mean a function g(x) =
(g
1
(x), . . . , g
n
(x)) such that g
1
, . . . , g
n
are in L
2
(R) and h(x, t) = g(x ct) is a solution
of (1.1), for some real number c. In the scalar case n = 1 (which includes the Korteweg-
de Vries equation as a specic example), it is well known that such waves often play an
important or even dominant role in the evolution of general solutions of nonlinear dispersive
wave equations (cf. [9]). This is due in large part to the remarkable stability properties of
solitary waves, which enable them to retain their identity even under large perturbations.
Theoretical explanations of the stability of solitary-wave solutions of (1.1) have undergone
active development in the past three decades, but has so far been restricted to the scalar
case (for a brief overview and some references, see [2]). It is our intention here to extend
some of this work to the case n > 1.
The approach we take to stability theory here is the same that has underlain all proofs
of stability of solitary waves (dating back to one given by Boussinesq himself in 1872 [12]).
First, we observe that equation (1.1) can be put in Hamiltonian form, and hence has the
Hamiltonian functional E itself as a conserved functional. Another conserved functional
Q is dened by Q(h) =
_

1
2
h, Dh) dx. It turns out (see Section 2 below) that g is
a solitary-wave solution of (1.1) if and only if g is a critical point for the constrained
variational problem of minimizing E over a level set of Q. Moreover, a standard argument
3
shows that if g is actually a local minimizer for E under this constraint, then G, the
intersection of the level sets of E and Q containing g, is a stable set of solitary waves.
This means that for every > 0, there exists > 0 such that if h is within of G (in an
appropriate norm) at time t = 0, then h remains within of G for all times t 0.
In Theorem 2.1 below, we give sucient conditions for the existence of stable sets of
solitary-wave solutions of (1.1). The conditions include one which is phrased in terms
of the above-mentioned variational problem, but as pointed out in Theorem 2.2, in some
important situations all the conditions can be reduced to simple properties of the func-
tion N and the symbol of the operator L. The proof of Theorem 2.1, which is given in
Section 3, proceeds by using P. Lions method of concentration compactness to show the
existence of a non-empty set of global minimizers of E on each level set of Q. The use of
concentration compactness to prove existence and stability of solitary waves goes back to
a paper of Cazenave and Lions on the nonlinear Schrodinger equation [13], and has since
been developed by a number of authors (see, e.g., [3,6,14,18,31]. Our point of departure is
the method of [2], which was easily adapted to handle the systems considered here.
In Section 4, we apply Theorems 2.1 and 2.2 to prove the existence of stable sets of
solitary-wave solutions to systems modelling the strong interaction of long internal waves
in stratied uids. In the rst of these systems, derived by Gear and Grimshaw in [20],
the components h
1
(x, t) and h
2
(x, t) of h(x, t) represent the slow horizontal variations, due
to weak nonlinearity and dispersion, of two long waves which in the linear, non-dispersive
limit correspond to two dierent vertical modes. In the other system, derived by Liu,
Kubota, and Ko in [27], h
1
and h
2
represent small, long-wavelength disturbances at two
pycnoclines separated by a region of constant density. (The question of how exactly the
situations governed by the two systems relate to each other physically is an interesting
one, to which the present authors do not yet know the answer. In particular, there is no
way to obtain one equation as a scaling limit of the other.)
For reasons mentioned earlier, the derivations of both systems assume that the waves
represented by h
1
and h
2
travel at nearly the same speed. It is also possible to derive
systems with n 3, describing the strong simultaneous interactions of three or more
underlying modes, and Theorem 2.1 applies to such systems as well. However, these
4
systems are of limited physical interest, since in a given uid it is relatively unlikely that
one can nd three linear modes whose corresponding wavespeeds are close enough for such
interactions to occur.
We note that an existence result for Liu-Kubota-Ko solitary waves appears in [3], and
an existence result for Gear-Grimshaw solitary waves appears in [10]. Both these papers
use the concentration compactness technique to obtain solitary waves as global minimizers
to constrained variational problems. However, since the minimized functional and the
constraint functional are not constants of the motion, these results do not yield the stability
of the solitary waves which are found to exist.
One motivation for the present study was provided by the numerical experiments con-
ducted in [20] and [27]. Interestingly, Liu, Kubota, and Ko did not observe anything
close to a steady travelling-wave solution of their system: instead, they found leap-frog
solutions in which localized disturbances in h
1
and h
2
took turns overtaking and falling
behind each other. Gear and Grimshaw, on the other hand, found that for typical val-
ues of the parameters in their equation, general initial data would quickly give rise to
steady travelling-wave solutions which maintained their identity even after colliding with
each other. They also were able to duplicate the leap-frogging behavior observed in [27]
by choosing their parameters so as to decouple the nonlinear terms in their system. The
stability results in the present paper validate the numerical observations of stable solitary
waves made by Gear and Grimshaw, and also show that the observed leap-frog solutions
do not arise due to lack of stability of solitary waves.
One issue which our stability result does not resolve, however, is that of the structure of
the stable sets of solitary waves. Indeed, this is a general drawback of the concentration-
compactness approach to stability as compared with other approaches involving ner anal-
ysis (cf. the discussion in [2]). This issue has bearing on the leap-frog solutions mentioned
in the preceding paragraph: if, for example, it were the case that the stable set of solitary
waves included functions g = (g
1
, g
2
) such that the maxima of g
1
and g
2
are located at
dierent points on the x-axis, then a leap-frog solution might actually represent a solution
which stays at all times very close to the stable set.
To shed light on this latter question, we investigate the symmetry properties of solitary-
5
wave solutions to the Liu-Kubota-Ko system in Section 5. In Theorem 5.4 we show that,
in case the coecients of the nonlinear terms in the system are positive, then the solitary
waves in the stable sets found in Section 4 must have components which are both symmetric
about the same value of x and which decay monotonically to zero away from their common
axis of symmetry. Hence, if a leap-frog solution exists in this case, it cannot be said to
closely resemble a solitary wave at any given time. This would suggest that while solitary
waves are stable in the sense of Theorem 2.1, they may not be asymptotically stable in the
sense of Lyapunov. (This would contrast with the strong asymptotic stability properties
of KdV solitary waves [30].) Unfortunately, there remains a gap in the evidence: since
the leap-frog solutions observed in [27] were for a system in which the coecients of the
nonlinear term were of mixed sign, it is not clear yet whether such solutions exist in the case
of positive coecients. On the other hand, we note that the leap-frog solutions observed
in [20] were obtained in the case in which the coecients of the nonlinear terms were both
positive.
Theorem 5.4 is actually closely related to a result of Maia [29] for the full equations
of motion of an incompressible, inviscid stratied uid. Maias work in turn represents a
development of the symmetry theory for solitary waves initiated by Craig and Sternberg
[16,17], in particular incorporating into the arguments of [17] some simplications suggested
by the work of Congming Li [25]. Our proof essentially follows the lines of Maias, with
some modications and further simplications appropriate to the present situation.
Finally, we also obtain, in Theorems 5.5 and 5.6, a monotonicity result for bore-like
solutions to the problem modeled by the Liu-Kubota-Ko system, and a symmetry result
for solitary-wave solutions of a scalar equation derived by Kubota, Ko, and Dobbs [24] as
a model for long internal waves in a stratied uid.
A preliminary version of Theorem 2.1 was announced in [4].
Notation. We use , ) to denote the usual inner product in C
n
; i.e., for v = (v
1
, . . . , v
n
)
and w = (w
1
, . . . w
n
) in C
n
we set v, w) = v
1
w
1
+ +v
n
w
n
, where bars denote complex
conjugation. For v in C
n
(or in R
n
) we dene [v[ = v, v)
1/2
.
Let I be an interval in R. As usual, for 1 p < , L
p
(I) denotes the set of all
measurable functions f : R R such that (
_

[f(x)[
p
dx)
1/p
< . We dene X
p
(I) to
6
be the Banach space of all measurable functions f : I R
n
such that [f[
X
p
(I)
< , where
[f[
X
p
(I)
=
__
I
[f(x)[
p
dx
_
1/p
.
For s R, let H
s
(I) denote the L
2
-based Sobolev space of order s on I, and dene
Y
s
(I) = (H
s
(I))
n
= f = (f
1
, . . . , f
n
) : f
i
H
s
(I) for i = 1, . . . , n, with norm given by
|f|
Y
s
(I)
= |f
1
|
H
s
(I)
+ +|f
n
|
H
s
(I)
.
In case I = R, the spaces X
p
(I) and Y
s
(I) will be denoted by X
p
and Y
s
, and the corre-
sponding norms will be denoted by [f[
p
and |f|
s
. We dene Y

to be the intersection of
all the spaces Y
s
as s ranges over the set of all real numbers.
If H is any Hilbert space then l
2
(H) will denote the Hilbert space of all innite sequences
x = (x
1
, x
2
, . . . ), x
i
H, such that
|x|
l
2
(H)
=
_
_

j=1
|x
j
|
2
H
_
_
1/2
< .
If is any open subset of R
n
, C
k
() denotes the set of all functions on whose partial
derivatives up to order k exist on , and C
k
(

) denotes the set of all functions whose

partial derivatives up to order k exist on and can be continuously extended to

. We
also dene C

() =

k=0
C
k
() and C

) =

k=0
C
k
(

).
If X is a Banach space and G is a subset of X, we say that a sequence x
n
in X
converges to G if
lim
n
inf
gG
|x
n
g|
X
= 0.
Also, for each T > 0, C([0, T]; X) will denote the Banach space of all continuous maps h
from [0, T] to X, with norm dened by |h|
C([0,T];X)
= sup
t[0,T]
|h(t)|
X
.
Hats will always denote Fourier transforms with respect to x:

(k) =
_

e
ikx
(x) dx,
where the integral is interpreted in the usual way for vector-valued functions .
2. Sucient conditions for stability of solitary waves.
Consider a vector-valued nonlinear dispersive wave equation of the form
h
t
+D
1
(N(h) Lh)
x
= 0, (2.1)
7
in which the unknown h is an R
n
-valued function of the variables x and t. The operators
D, N, and L in (2.1) are dened as follows:
D is an n n diagonal matrix with positive entries
i
along the diagonal.
N is the gradient of a function N : R
n
R. We assume that N is homogeneous of
degree p + 2, where p is any positive number; or, in other words,
N(v) =
p+2
N(v)
for every v R
n
and every > 0. Further, we require N to be twice continuously
dierentiable on the unit sphere in R
n
(and hence everywhere on R
n
). In particular, it
follows from our assumptions on N that

N(f) dx

C[f[
p+2
p+2
for all f X
p+2
, where C is independent of f: to see this, notice that [N(f)[ =
[f[
p+2
N(f/[f[) C[f[
p+2
, where C is the supremum of N on .
The dispersion operator L is a matrix Fourier multiplier operator dened by

Lh(k) = A(k)

h(k)
for k R, where A(k), the symbol of L, is for each k R a symmetric n n matrix with
real entries, and A(k) satises A(k) = A(k) for all k R.
Further, we make the following assumptions on A(k):
(A1) There exist positive constants C
1
, C
2
and a number s > p/4 such that
C
1
[k[
2s
[v[
2
A(k)v, v) C
2
[k[
2s
[v[
2
for all vectors v in R
n
and all suciently large values of [k[.
(A2) For each i and j between 1 and n, the matrix components a
ij
(k) are four times
dierentiable on k ,= 0. Moreover, there exist constants C and K such that for all
m 1, 2, 3, 4,

_
d
dk
_
m
_
a
ij
(k) a
ij
(0)
k
_

C[k[
m
for 0 < [k[ K,
8
and

_
d
dk
_
m
_
_
[a
ij
(k)[
k
s
_

C[k[
m
for [k[ K.
We remark that the condition in assumption (A2) on the behavior of a
ij
near the origin
is satised whenever the derivatives up to order ve of a
ij
(k) exist and are bounded on
(0, K].
A somewhat stronger condition on A(k), which implies both (A1) and (A2) and has the
advantage of being more convenient to verify, is the following:
(A3) The symmetric matrix A(k) has n distinct eigenvalues
1
(k), . . . ,
n
(k) which, to-
gether with their derivatives up to order ve, are bounded on 0 < k < 1 and contin-
uous on 0 < k < . Furthermore, there exist positive constants C
1
, C
2
, and K such
that for 1 i n and 0 m 4 one has
C
1
[k[
2sm

_
d
dk
_
m

i
(k) C
2
[k[
2sm
for [k[ > K.
That (A3) implies (A1) and (A2) follows from the perturbation theory expounded in
chapter II of [23] (see in particular Section II.5.3). Note also that if the
i
(k) are assumed
to be analytic functions of k for k > 0, then the assumption that the eigenvalues are
distinct may be dropped (cf. Theorem II.1.10 of [23]).
We will assume in what follows that equation (2.1) is globally well-posed in Y
r
for some
r s. In other words, we assume that for every h
0
Y
r
and every T > 0, there exists
a unique weak solution h of (2.1) in C([0, T]; Y
r
), and the correspondence h
0
h denes
a continuous map from Y
r
to C([0, T]; Y
r
). Here weak solution means any element h of
C([0, T]; Y
r
) such that for all t 0, h
t
exists in Y
r
(in the usual sense of the derivative
of a Banach-space valued function), and is equal to D
1
(N(h) Lh)
x
. Notice that
our assumption on N guarantees that, for xed t, N(h(t)) is in L
2/(p+1)
(R), and hence
that D
1
(N(h) Lh)
x
exists as a tempered distribution on R, so that the equality has
sense.
In particular, we are concerned with solitary-wave solutions of (2.1), which by denition
are solutions of the form h(t) = ( ct), where Y
r
and c is a real number called the
9
wavespeed of the solitary wave. We also refer to the prole itself as a solitary wave. Thus
Y
r
is a solitary wave if and only if it satises the equation
cD = L N(). (2.2)
We now dene functionals Q and E on Y
s
which are constants of the motion for (2.1)
and which play a crucial role in the stability theory for solitary-wave solutions. Let
Q(f) =
_

1
2
f, Df) dx
and
E(f) =
_

1
2
f, Lf) N(f) dx.
We claim that if h is a solution of (2.1) in C([0, T]; Y
r
) then Q(h(x, t)) and E(h(x, t)) are
independent of t. Indeed, taking the inner product of (2.1) with Dh and integrating over
R, one sees that
d
dt
Q(h(x, t)) = 0, at least if h is in C([0, T]; Y
r
) for r

suciently large.
Hence Q(h(x, t)) = Q(h(x, 0)) for all t if h is a solution in C([0, T]; Y
r
), and the result
for solutions h in C([0, T]; Y
r
) then follows from the assumed well-posedness properties
of (2.1) and the fact that Y
r
is dense in Y
r
. Next, observe that (2.1) may be written in
Hamiltonian form as
h
t
= J E(h),
where E denotes the Frechet derivative of E and
J =
x
D
1
is antisymmetric with respect to the inner product in Y
r
. It follows that E plays the role
of a Hamiltonian functional for (2.1), and in particular is a constant of the motion.
The importance of the functionals E and Q for our purposes rests on the fact that (2.2)
can be written in the form
E() = c Q(). (2.3)
We will show that, under the assumptions stated below in Theorem 2.1, the problem of
minimizing E subject to constant Q always has a non-empty solution set. But since each
10
element of the solution set must satisfy the Euler-Lagrange equation (2.3), the solution set
must consist of solitary waves.
Actually, in what follows it will be more convenient to work with a modied functional
E
0
than with the functional E dened above. To dene E
0
, rst consider the operator
D+L, where ranges over the set of real numbers. From the perturbation theory of sym-
metric matrices (see Theorem II.6.8, p. 122 of [23]), it follows that there exist n functions

1
(k, ), . . . ,
n
(k, ), representing the (unordered, and possibly repeated) eigenvalues of
D+A(k), which depend smoothly on and, for a given , have the same dierentiability
and continuity properties with respect to k as do the functions a
ij
(k).
From the variational characterization of eigenvalues, we have that the least eigenvalue
of D+A(k) is the inmum of the set of values of (D+A(k))v, v) as v ranges over the
set of vectors in R
n
such that |v| = 1. It follows easily that the function b() dened by
b() = inf
i
(k, ) : 0 k < and 1 i n
is a strictly decreasing function of . Moreover, since
b()
_
min
1in

i
_
+b(0),
and b(0) > as a consequence of (A1) and (A2), then b() > 0 for suciently large.
Also, since for any given k one has
b() sup
v=1
(D +A(k))v, v)
_
max
1in

i
_
+ max
1in

i
(k, 0),
it follows that b() < 0 for suciently large and negative. We conclude that there exists
a unique
0
such that b(
0
) = 0.
The number
0
can be characterized as the smallest value of such that the matrix
D + A(k) is non-negative for all k R. Alternatively, we can view
0
as the greatest
possible eigenvalue of D
1
A(k), as k ranges over R. Hence
0
is the largest possible
wavespeed of innitesimal sinusoidal waves, i.e.,
0
is the largest value of such that the
linearized equation
h
t
D
1
(Lh)
x
= 0
has a solution of the form h(x, t) = ve
ik(xt)
with nonzero v R
n
.
11
We now dene =
0
D +L, and dene the functional E
0
by
E
0
(f) =
_

1
2
f, f) N(f) dx,
so that E
0
=
0
Q + E. Notice that replacing L by
0
D + L in (2.1) amounts to nothing
more than changing to new coordinates x

and t

given by x

= x
0
t and t

= t. Thus,
up to a Galilean coordinate change, one can always assume that = L and E
0
= E.
Dene the number I
q
by
I
q
= infE
0
(f) : f Y
s
and Q(f) = q
The set of minimizers for I
q
is
G
q
= g Y
s
: E
0
(g) = I
q
and Q(g) = q,
and the Euler-Lagrange equation for the constrained minimization problem solved by the
functions in G
q
is
E
0
(g) = E(g) +
0
Q(g) = Q(g),
where is the Lagrange multiplier. Comparing this equation with (2.3), we see that if
g G
q
, then g is a solitary-wave solution of (2.1) with wavespeed c =
0
. (Notice that
the multiplier could, in principle, vary from one element of G
q
to the next).
We can now state the following result, giving a sucient condition for the existence of
a stable set of solitary-wave solutions of (2.1).
Theorem 2.1. Suppose that s, p, and L are such that (A1) and (A2) hold. If the solution
I
q
of the variational problem dened above satises I
q
< 0 for all q > 0, then for each
q > 0 the set G
q
of minimizers for the variational problem is non-empty, and each g G
q
is a solitary-wave solution of (2.1) with wavespeed c >
0
. Moreover, G
q
is a stable set
of initial data for (2.1), in the following sense: for every > 0 there exists such that if
h
0
Y
r
and
inf
gG
q
|h
0
g|
s
< ,
then the solution h(x, t) of (2.1) with h(x, 0) = h
0
satises
inf
gG
q
|h(x, t) g|
s
<
12
for all t R.
The proof of Theorem 2.1 is given in Section 3 below.
The next result, which is a corollary of Theorem 2.1, will apply to the model equations
considered in Section 4.
Theorem 2.2. Suppose that s, p, and L are such that (A1) and (A2) hold. Suppose also
that there exists a vector v
0
R
n
such that N(v
0
) > 0 and
[v
0
, (
0
D +A(k))v
0
)[ C[k[
s
0
for all [k[ 1, (2.4)
where C > 0 and s
0
> p/2. Then for each q > 0, the set G
q
is non-empty and the elements
g of G
q
are solitary waves with wavespeeds c greater than
0
. Moreover, G
q
is stable in
the sense of Theorem 2.1.
Remarks.
(i) In particular, inequality (2.4) holds in the important special case when
0
D + A(k)
has the eigenvalue 0 at k = 0. This may be seen by taking v
0
to be an eigenvector
for the eigenvalue 0 of
0
D + A(0); then v
0
, (
0
D + A(k))v
0
) denes a function of
k which has the value 0 at k = 0 and has bounded derivative on 0 k 1, and it
follows that (2.4) holds for s
0
at least 1.
(ii) If N(v) = N(v) for v R
n
, then we can drop the condition that N(v
0
) > 0, since
v
0
may be replaced by v
0
if necessary.
Proof. We claim that the existence of a vector v
0
with the stated properties implies that
I
q
< 0 for each q > 0. To see this, let w(x) = v
0
(x), where (x) is any non-negative
smooth function with compact support, normalized so that Q(w) = q. For any > 0 let
w

(x) =

w(x). Then by assumption there exists a constant C such that for [k[ 1
and < 1/K, where K is as in (A2),

, w

) dx

=
1

v
0
, (
0
D +A(k))v
0
)[

(k/)[
2
dk
=
_

v
0
, (
0
D +A(k))v
0
)[

(k)[
2
dk
C
s
0
_
|k|1/
[k[
s
0
[

(k)[
2
dk +C
2s
_
|k|1/
[k[
2s
[

(k)[
2
dk.
13
But because is smooth with compact support,

(k) decays more rapidly than any power
of k as [k[ , and it follows that the last integral in the preceding expression vanishes
more rapidly than any power of as 0. Therefore

, w

) dx

C
s
0
,
for all small values of , where C is independent of .
On the other hand,
_

N(w

) dx =
p/2
_

N(w) dx
=
p/2
N
_
v
0
[v
0
[
_
[v
0
[
p+2
_

(x)
p+2
dx = C
p/2
,
where C > 0 is independent of .
We conclude that in the expression
E
0
(w

) =
_

1
2
w

, w

) dx
_

N(w

) dx
the second integral on the right-hand side is positive and (since s
0
> p/2) goes to zero
more slowly than the rst term as 0. It follows that E
0
(w

) < 0 for suciently near

zero. On the other hand, one has Q(w

) = q for all . Therefore I

q
must be less than zero,
as claimed.
The conclusion of Theorem 2.2 now follows from Theorem 2.1.
3. Proof of Theorem 2.1.
The proof of Theorem 2.1 proceeds via P. Lions method of concentration compactness
[26], and follows the lines of the proof of stability of ground-state solutions of the nonlinear
Schrodinger equation given by Cazenave and Lions in [13].
We begin with the following standard estimate.
Lemma 3.1. Suppose I is an interval in R, p > 0, and s > p/4. Then there exists C > 0
such that for all f Y
s
(I),
[f[
p+2
X
p+2
(I)
C|f|
p/2s
Y
s
(I)
[f[
p+2(p/2s)
X
2
(I)
.
14
Proof. Let s

=
p
2(p+2)
. From the Sobolev embedding theorem, it follows that there exists
a constant C independent of f such that for all f Y
s
(I),
[f[
X
p+2
(I)
C|f|
Y
s
(I)
.
The stated result then follows from the interpolation inequality
|f|
Y
s
(I)
C|f|
s

/s
Y
s
(I)
|f|
1(s

/s)
Y
0
(I)
,
since Y
0
(I) = X
2
(I).
Lemma 3.2. For all q > 0, we have I
q
> .
Proof. Let f be an arbitrary element of Y
s
satisfying Q(f) = q; we wish to show that
E
0
(f) is bounded below by a number which is independent of f.
From assumption (A1) and the denition of
0
, it follows that there exist constants
C
3
> 0 and C
4
> 0 such that
C
3
(1 +[k[)
2s
[v[
2
v, [(
0
+ 1)D +A(k)] v) C
4
(1 +[k[)
2s
[v[
2
for all k R and v C
2
. Therefore the expression
__

1
2
f(x), f(x)) dx +Q(f)
_
1/2
denes a norm on Y
s
equivalent to |f|
s
. In particular, it follows that we can write
E
0
(f) = E
0
(f) +Q(f) Q(f)
=
_

1
2
f(x), f(x)) dx +Q(f)
_

N(f) dx Q(f)
C
3
|f|
2
s
C[f[
p+2
p+2
q,
where C is a positive constant which is independent of f. But by Lemma 3.1 and Youngs
Inequality,
[f[
p+2
p+2
C|f|
p/2s
s
[f[
p+2(p/2s)
2
|f|
2
s
+C[f[
2+4sp/(4sp)
2
,
where > 0 can be chosen arbitrarily small, and again C denotes various constants which
may depend on but do not depend on f. Combining with the preceding estimate, and
taking < C
3
, we now obtain
E
0
(f) C[f[
2+4sp/(4sp)
2
q.
15
The proof concludes with the observation that [f[
2
is dominated by a constant times Q(f),
and hence remains bounded due to the assumption that Q(f) = q.
We dene a minimizing sequence for I
q
to be any sequence f
n
of functions in Y
s
satisfying
Q(f
n
) = q for all n
and
lim
n
E
0
(f
n
) = I
q
.
To each minimizing sequence f
n
is associated a sequence of nondecreasing functions
M
n
: [0, ) [0, q] dened by
M
n
(r) = sup
yR
_
y+r
yr
1
2
f
n
, Df
n
) dx.
A standard argument shows that any uniformly bounded sequence of nondecreasing func-
tions on [0, ) must have a subsequence which converges pointwise to a nondecreasing
limit function on [0, ). Hence M
n
has such a subsequence, which we denote again by
M
n
. Let M : [0, ) [0, q] be the nondecreasing function to which M
n
converges, and
dene
= lim
r
M(r),
so 0 q.
Lemma 3.3. If f
n
is a minimizing sequence for I
q
, then there exist constants B > 0
and
2
> 0 such that
(i) |f
n
|
s
B for all n and
(ii)
_

N(f
n
) dx
2
for all suciently large n.
Proof. As was noted in the proof of Lemma 3.2, the quantity
__

1
2
f(x), f(x)) dx +Q(f)
_
1/2
denes a norm on Y
s
equivalent to |f|
s
. Therefore
|f
n
|
2
s
C
__

1
2
f
n
(x), f
n
(x)) dx +Q(f
n
)
_
C
_
sup
n
E
0
(f
n
) +[f
n
[
p+2
p+2
+q
_
C
_
1 +[f
n
[
p+2(p/2s)
2
|f
n
|
p/2s
s
_
.
16
where Lemma 3.1 has been used, and C denotes constants which are independent of f Y
s
.
But since Q(f
n
) = q for all n, then [f
n
[
2
remains bounded and so we have
|f
n
|
2
s
C
_
1 +|f
n
|
p/2s
s
_
.
Since p/2s < 2, the existence of the bound B follows immediately.
To prove (ii), suppose that such a constant
2
does not exist. Then
liminf
n
_

N(f
n
) dx 0.
Also, from the denition of it follows that
_

f
n
(x), f
n
(x)) dx 0
for all n. Hence
I
q
= lim
n
E
0
(f
n
)
= lim
n
__

1
2
f
n
(x), f
n
(x)) dx
_

N(f
n
) dx
_
liminf
n
_

N(f
n
) dx 0,
which contradicts the assumption that I
q
< 0.
Lemma 3.4. For all q
1
, q
2
> 0, one has
I
q
1
+q
2
< I
q
1
+I
q
2
.
Proof. First we claim that for > 1 and q > 0,
I
q
< I
q
.
In fact, let f
n
be a minimizing sequence for I
q
, and notice that for all n, Q(

f
n
) = q
and hence E
0
(

f
n
) I
q
. It follows that
I
q
E
0
(

f
n
) = E
0
(f
n
) + (
(p+2)/2
)
_

N(f
n
) dx;
17
and taking n and using Lemma 3.3(ii), we conclude that
I
q
I
q
+ (
(p+2)/2
)
2
< I
q
,
as claimed.
Now in case q
1
> q
2
, then from what was just shown it follows that
I
(q
1
+q
2
)
= I
q
1
(1+q
2
/q
1
)
< (1 +
q
2
q
1
)I
q
1
< I
q
1
+
q
2
q
1
_
q
1
q
2
I
q
2
_
= I
q
1
+I
q
2
;
whereas in the case q
1
= q
2
we have
I
(q
1
+q
2
)
= I
2q
1
< 2I
q
1
= I
q
1
+I
q
2
.
The next step in the proof of Theorem 2.1 is to rule out the possibilities that 0 < < q
and that = 0. The former of these two possibilities is dealt with in the next three
lemmas, which represent a simplication and generalization of an argument appearing in
Section 4 of [2].
Lemma 3.5. Let
P =
_
s
2
_
+ 1,
where the brackets denote the greatest integer function. We can write =
1
+ (
2
)
2
,
where
1
and
2
are self-adjoint operators on Y
s
with the following properties:
(i) There exists a constant C > 0 such that if is any function which is in L

(R)
and has derivative

in L

(R), and f is any function in X

2
, then
[[
1
, ]f[
2
C[

[f[
2
,
where [
1
, ] denotes the commutator
1
(f) (
1
f).
(ii) There exists a constant C > 0 such that if is any function which is in L

(R)
and has derivatives up to order P in L

(R), and f is any function in X

2
, then
[[
2
, ]f[
2
C
_
P

i=1

d
i

dx
i

_
[f[
2
.
18
Proof. First choose a function (k) C

0
(R) such that (k) = 1 for [k[ < K, where K is
the constant dened in assumption (A2) above. Dene A
1
(k) = (k)(
0
D + A(k)), and
dene A
2
(k) to be the square root of the positive denite matrix (1 (k))(
0
D+A(k)).
Since
0
D+A(k) = A
1
(k) +(A
2
(k))
2
, then =
1
+
2
2
, where
1
and
2
are the Fourier
multiplier operators with symbols A
1
(k) and A
2
(k).
Now, for given values of i and j between 1 and n, let (a
1
)
ij
(k) be the entry in the ith
row and jth column of A
1
(k), and let (
1
)
ij
denote the scalar Fourier multiplier operator
with symbol (a
1
)
ij
(k). Let

= (
1
)
ij
(a
1
)
ij
(0); then we can write

=
d
dx
T where T is
the operator with symbol
(k) =
(a
1
)
ij
(k) (a
1
)
ij
(0)
k
.
By assumption (A2), we have that sup
kR
[k[
m

_
d
dk
_
m
(k)

< for 0 m 4; and it

then follows from Theorem 35 of [15] that
[[T, ]f

[
2
C[

[f[
2
,
for some C independent of f L
2
(R) and (As stated in [15], Theorem 35 actually
requires estimates on for all m 0, but the proof given there shows that it suces to
have estimates for 0 m 4.) Since
[[(
1
)
ij
, ]f[
2
=

, ]f

2
=

T
d
dx
(f) Tf

2
[T(

f)[
2
+[[T, ]f

[
2
,
and T is bounded on L
2
, it follows that
[[(
1
)
ij
, ]f[
2
C[

[f[
2
for all f L
2
. Finally, since for f = (f
1
, . . . , f
n
) X
2
one has
[[
1
, ]f[
2
=

i,j=1
[(
1
)
ij
, ]f
j

i,j=1
[[(
1
)
ij
, ]f
j
[
2
,
it follows that (i) holds for
1
.
Similarly, to prove (ii) it suces to verify that the same statement holds for all f L
2
(R)
if
2
is replaced by its ijth entry (
2
)
ij
. But this is exactly the content of part 2 of Lemma
4.2 of [2], since (
2
)
ij
has the same properties as the operator M
2
dened there.
19
Lemma 3.6. For every > 0, there exist a number N N and sequences g
N
, g
N+1
,
and h
N
, h
N+1
, of functions in Y
s
such that for every n N,
(i) [Q(g
n
) [ < ,
(ii) [Q(h
n
) (q )[ < , and
(iii) E
0
(f
n
) E
0
(g
n
) +E
0
(h
n
) .
Proof. Choose C

0
with support in [2, 2] such that 1 on [1, 1], and let C

be such that
2
+
2
1 on R. For each r R dene
r
(x) = (x/r) and
r
(x) = (x/r).
From the denition of it follows that for every suciently large value of r, one can
nd N = N(r) such that for all n N,
< M
n
(r) M
n
(2r) < +.
In particular, we can nd y
n
such that
_
y
n
+r
y
n
r
1
2
f, Df) dx >
and
_
y
n
+2r
y
n
2r
1
2
f, Df) dx < +.
It follows that if we dene g
n
(x) =
r
(x y
n
)f
n
(x) and h
n
(x) =
r
(x y
n
)f
n
(x), then
(i) and (ii) hold for all n N(r). We now show that if r is chosen suciently large, then
(iii) also holds for all such n, if in (iii) is replaced by C

for certain positive numbers C

and .
Begin by writing
E
0
(g
n
) =
1
2
__

g
n
,
1
g
n
) dx +
_

2
g
n
,
2
g
n
) dx
_

N(g
n
) dx. (3.1)
The rst of the integrals on the right-hand side of (3.1) can be written as
_

r
f
n
,
1
(
r
f
n
)) dx =
_

2
r
f
n
,
1
f
n
) dx +
_

r
f
n
, [
1
,
r
]f
n
) dx.
Now by Lemma 3.5 (i),

r
f
n
, [
1
,
r
]f
n
) dx

[
r
f
n
[
2
[[
1
,
r
]f
n
[
2
C[

r
[

[f
n
[
2
2
.
20
But since [

r
[

= [

/r, and [f
n
[
2
is bounded independently of n, it follows that

r
f
n
, [L,
r
]f
n
) dx

C/r
where the constant C is independent of r, n and .
Similarly, writing the second integral on the right-hand side of (3.1) as
_

2
r

2
f
n
,
2
f
n
) dx + 2
_

2
f
n
, [
2
,
r
]f
n
) dx +[[
2
,
r
]f
n
[
2
2
,
and using Lemma 3.5 (ii) and the fact that
2
is a bounded operator from Y
s
to X
2
, we
see that
_

2
g
n
,
2
g
n
) dx
_

2
r

2
f
n
,
2
f
n
) dx +C/r,
where again C is independent of r, n and .
Finally, since

N(g
n
)
2
r
N(f
n
) dx

(
p+2
r

2
r
)N(f
n
) dx

C([f
n
[
X
p+2
(I
1
)
+[f
n
[
X
p+2
(I
1
)
)
p+2
,
where I
1
and I
2
denote the intervals [y
n
2r, y
n
r] and [y
n
+r, y
n
+ 2r], it follows from
Lemma 3.1 that

N(g
n
)
2
r
N(f
n
) dx

C|f
n
|
p/2s
s
_
[f
n
[
X
2
(I
1
)
+[f
n
[
X
2
(I
2
)
_
p+2(p/2s)
C

,
where = p + 2 (p/2s) and C is independent of r, n, and .
Substituting these inequalities in (3.1) yields
E
0
(g
n
)
_

2
r
_
1
2
f
n
, f
n
) N(f
n
)
_
dx +C(1/r +

).
The same argument yields the result
E
0
(h
n
)
_

2
r
_
1
2
f
n
, f
n
) N(f
n
)
_
dx +C(1/r +

),
and it follows that
E
0
(g
n
) +E
0
(h
n
) E(f
n
) +C(1/r +

).
21
Choosing r 1/

, we conclude that there exists a constant C, independent of , such

that
E(f
n
) E(g
n
) +E(h
n
) C

for all n N(r).

This proves the Lemma, except that (iii) has been modied by replacing by C

. But
since C and are independent of , we can now apply what has just been proved to ,
where is chosen to be less than the minimum of and (/C)
1/
; it follows that the Lemma
holds as stated.
Lemma 3.7. If 0 < < q then
I
q
I

+I
q
.
Proof. First, we claim that if is any real number and f Y with |f|
Y
B and
[Q(f) [ /2, then
I

E
0
(f) +C[Q(f) [,
where C depends only on and B. To see this, let

f =

f where = /Q(f). Then
Q(

f) = , and so
I

E
0
(

f) = E
0
(f) + ( 1)E
0
(f) +(1
p/2
)
_

N(f) dx
E
0
(f) +C
_
[1 [ +[1
p/2
[
_
.
But [Q(f) [ /2 implies that 2 and that [1
p/2
[ < C[1 [ < C[Q(f) [, so
the claim has been proved.
The preceding observation together with Lemma 3.6 implies that there exists a sub-
sequence f
n
k
of f
n
and corresponding functions g
n
k
and h
n
k
such that for all
k,
E
0
(g
n
k
) I

1
k
E
0
(h
n
k
) I
q

1
k
E
0
(f
n
k
) E
0
(g
n
k
) +E
0
(h
n
k
)
1
k
.
22
Thus
E
0
(f
n
k
) I

+I
q

3
k
,
and the desired result follows by taking the limit of both sides as k .
The next two lemmas are used to dispose of the possibility that = 0.
Lemma 3.8. Suppose B > 0 and > 0 are given. Then there exists
1
=
1
(B, ) > 0
such that if f Y
s
with |f|
s
B and [f[
p+2
, then
[f[
X
p+2
(I)

1
for some interval I R of length 4.
Proof. Choose : R [0, 1] smooth with support in [2, 2] and satisfying

jZ
(xj) = 1
for all x R, and dene
j
= (x j) for all j Z. The map T : Y
s
l
2
(Y
s
) dened by
Tw =
j
w
jZ
is bounded (this is clear in the case when s is a non-negative integer, and the case for
general s 0 then follows by interpolation: see, e.g., Section 5.6 of [8]). Therefore we can
nd C
0
> 0 such that

jZ
|
j
f|
2
s
C
0
|f|
2
s
for all f Y
s
.
Now let C
1
be a positive number such that

jZ
[(xj)[
3
C
1
for all x R, and dene
C
2
=
C
0
B
2
C
1
. We claim that for every nonzero f Y
s
there exists j
0
Z such that
|
j
0
f|
2
s
(1 +C
2
[f[
(p+2)
p+2
)[
j
0
f[
p+2
p+2
.
In fact, if no such j
0
exists, then one has
|
j
f|
2
s
> (1 +C
2
[f[
(p+2)
p+2
)[
j
f[
p+2
p+2
for every j Z. But then summing over j leads to
C
0
B
2
> (1 +C
2
[f[
(p+2)
p+2
)C
1
[f[
p+2
p+2
= C
1
[f[
p+2
p+2
+C
0
B
2
,
23
which is a contradiction.
Since [f[
p+2
, it follows from our claim that
|
j
0
f|
2
s
(1 +C
2

(p+2)
)[
j
0
f[
p+2
p+2
.
Now since s > p/4 > p/(2p + 4), from Sobolevs embedding theorem it follows that
[
j
0
f[
p+2
[f[
p+2
C
3
|f|
s
,
with C
3
independent of f. Therefore
[
j
0
f[
p+2
(C
2
3
(1 +C
2

(p+2)
))
1/p
,
and hence the Lemma has been proved, with
1
= (C
2
3
(1 + C
2

(p+2)
))
1/p
and I =
[j
0
2, j
0
+ 2].
Lemma 3.9. For every minimizing sequence f
n
, we have > 0.
Proof. From Lemmas 3.3 and 3.8 we deduce that there exist
1
> 0 and a sequence of
intervals I
n
= [y
n
2, y
n
+ 2] such that
[f
n
[
p+2
X
p+2
(I
n
)

1
for all suciently large n. Then Lemma 3.1, together with Lemma 3.3(i), gives

1
CB
p/2s
[f
n
[

X
2
(I
n
)
C
__
y
n
+2
y
n
2
f
n
, Df
n
) dx
_
/2
for all suciently large n, where = p + 2 (p/2s) and C is independent of n. Hence
= lim
r
M(r) M(2) = lim
n
M
n
(2)
1
2
_

1
C
_
2/
> 0.
Note now that Lemmas 3.4, 3.7, and 3.9 combine to show that = q. Therefore we can
apply the following result:
Lemma 3.10. Suppose = q. Then there exists a sequence of real numbers y
1
, y
2
,
such that the sequence

f
n
dened by

f
n
(x) = f
n
(x +y
n
) for all x R
has a subsequence converging in Y
s
norm to a function g G
q
.
We omit the proof of Lemma 3.10, since it diers in only minor details from the proof
of Lemma 2.5 of [2]; and the modications which are required are obvious.
24
Lemma 3.11. The set G
q
is not empty. Moreover, if f
n
is any minimizing sequence
for I
q
, then
(i) there exists a sequence y
1
, y
2
, ... and an element g G
q
such that f
n
( + y
n
)
has a subsequence converging strongly in Y
s
to g.
(ii)
lim
n
inf
gG
q
yR
|f
n
( +y) g|
s
= 0.
(iii) f
n
converges to G
q
in Y
s
.
The same conclusions hold for f
n
under the weaker hypothesis that Q(f
n
) q and
E
0
(f
n
) I
q
as n .
Proof. Lemmas 3.4, 3.7 and 3.9 show that = q; it then follows from Lemma 3.10 that G
q
is nonempty and that (i) holds for any minimizing sequence f
n
. If, on the other hand,
we assume only that Q(f
n
) q as n , then we still can assert that (i) holds for the
minimizing sequence
n
f
n
, where
n
=
_
q/Q(f
n
). But since
n
1, the convergence of
a subsequence of
n
f
n
( +y
n
) to g in Y
s
implies the convergence of the same subsequence
of f
n
( +y
n
) to g. Thus (i) holds under the weaker hypothesis on f
n
.
To complete the proof it suces to show that (i) implies (ii) and (iii). But (ii) follows
immediately from (i) and the fact that every subsequence of a minimizing sequence is
itself a minimizing sequence; and (iii) follows from (ii) and the fact that the functionals
E
0
and Q (and hence also the set G
q
) are invariant under the operation of replacing f by
f( +y).
We can now complete the proof of Theorem 2.1. It has already been shown in Lemma
3.11 that G
q
is non-empty. It remains therefore to show that the solitary waves in G
q
have
wavespeeds greater than
0
, and that the set G
q
is stable.
It follows from the denition of G
q
and the Lagrange multiplier principle (cf. Theorem
7.7.2 of [28]) that for each g G
q
there exists R such that
E
0
(g) = Q(g),
25
where the Frechet derivatives E
0
(g) and Q(g) are given by
E
0
(g) = g N(g) =
0
Dg +Lg N(g),
Q(g) = Dg.
Hence g solves (2.2) with c =
0
; i.e., the wavespeed of the solitary wave g is
0
.
We wish to show that < 0.
Note rst that
d
d
[E
0
(g)]
=1
=
d
d
_

2
_

1
2
g, g) dx
p+2
_

N(g) dx
_
=1
=
_

g, g) dx (p + 2)
_

N(g) dx
= 2E
0
(g) p
_

N(g) dx.
But E
0
(g) = I
q
< 0, and
_

N(g) dx > 0 by Lemma 3.3(ii), so

d
d
[E
0
(g)]
=1
< 0.
On the other hand, from the denition of the Frechet derivative we have
d
d
[E
0
(g)]
=1
=
_

E
0
(g),
d
d
[g]
=1
) dx
=
_

Q(g), g) dx =
_

g, Dg) dx;
and since
_

g, Dg) dx > 0 it follows that < 0 as claimed.

Now suppose that G
q
is not stable. Then there exists a sequence of solutions h
n
of
(2.1) and a sequence of times t
n
such that h
n
(, 0) converges to G
q
in Y
s
, but h
n
(, t
n
)
does not converge to G
q
in Y
s
. Since E
0
and Q are constants of the motion for (2.1) and are
continuous on Y
s
, it follows that Q(h
n
(, t
n
)) q and E
0
(h
n
(, t
n
)) I
q
as n . Hence
from Lemma 9(iii) it follows that h
n
(, t
n
) converges to G
q
in Y
s
, a contradiction.
4. Applications to model systems for long waves.
a) The Gear-Grimshaw system.
The Gear-Grimshaw system was derived in [20] to model the strong interaction of two
long internal gravity waves in a stratied uid, where the two waves are assumed to
26
correspond to dierent modes of the linearized equations of motion. Following [11], we
write it as
h
1t
+h
1
h
1x
+a
1
h
2
h
2x
+a
2
(h
1
h
2
)
x
+h
1xxx
+a
3
h
2xxx
= 0
b
1
h
2t
+rh
2x
+h
2
h
2x
+b
2
a
2
h
1
h
1x
+b
2
a
1
(h
1
h
2
)
x
+b
2
a
3
h
1xxx
+h
2xxx
= 0,
(4.1)
where a
1
, a
2
, a
3
, b
1
, b
2
, and r are real constants with b
1
, b
2
positive.
The system (4.1) can be rewritten in the form (2.1) by putting
N(h
1
, h
2
) =
1
2
_
h
3
1
3
+a
1
h
1
h
2
2
+a
2
h
2
1
h
2
+
b
1
2
h
3
2
3
_
and
D =
_
1 0
0 b
1
b
1
2
_
;
and dening the symbol A(k) of L by
A(k) =
_
k
2
a
3
k
2
a
3
k
2
b
1
2
k
2
_
= k
2
_
1 a
3
a
3
b
1
2
_
.
We verify that (4.1) satises the assumptions required by the stability theory of Section
2. First, in [11], it was shown that if b
2
a
2
3
< 1, then (4.1) is globally well-posed in Y
r
for
every r 1. In the case of (4.1), the function N appearing in (2.1) is homogeneous of
degree 3, so we take p = 1. The signs of the eigenvalues of A(k) are independent of k, and
both are positive if and only if b
2
a
2
3
< 1, so (A3) holds in this case with s = 1. One sees
easily that then
0
= 0. From the formula for N, we see that no matter what the values
of the parameters a
i
and b
i
, one can always nd v
0
R
2
such that N(v
0
) > 0, and (2.4)
obviously holds for any v
0
R
2
, with s
0
= 2. Hence from Theorem 2.2 we obtain the
following result.
Theorem 4.1. Suppose that b
2
a
2
3
< 1. Let E and Q be the invariant functionals associated
with (4.1), as dened in Section 2. Then for each q > 0, the problem of minimizing E
subject to the constraint Q = q has a nonempty solution set G
q
, and for each g G
q
, there
exists c > 0 such that g(x ct) is a solution of (4.1). Moreover, the set G
q
is stable in
the sense that for every > 0, there exists > 0 with the following property: if h
0
is any
function in Y
1
satisfying
|h
0
g|
1
<
27
for some g G
q
, then there exists a global solution h(x, t) of (4.1) with h(x, 0) = h
0
and
a map t g(t) from [0, ) to G
q
such that
|h(, t) g(t)|
1
<
for all t 0.
b) The Liu-Kubota-Ko system.
The Liu, Kubota & Ko system was derived in [27] to model the interaction between
a disturbance h
1
(x, t) located at an upper pycnocline and another disturbance h
2
(x, t)
located at a lower pycnocline in a three-layer uid. It can be written as
h
1t
c
1
h
1x
+
1
h
1
h
1x

1
(M
1
h
1
)
x

2
[(M
2
h
1
)
x
(Sh
2
)
x
] = 0
h
2t
c
2
h
2x
+
2
h
2
h
2x

3
(M
3
h
2
)
x

4
[(M
2
h
2
)
x
(Sh
1
)
x
] = 0. (4.2)
Here c
1
, c
2
,
1
,
2
,
1
,
2
,
3
,
4
are real constants, with
i
positive for i = 1, 2, 3, 4. The
operators M
1
, M
2
, M
3
are Fourier multiplier operators dened for H
1/2
by

M
i
(k) = m
i
(k)

(k),
where
m
i
(k) = k coth(kH
i
)
1
H
i
for i = 1, 2, 3; with H
1
, H
2
, H
3
being positive constants related to the depths of the three
uid layers. The operator S is also a Fourier multiplier operator,

S(k) = n(k)

(k),
where
n(k) =
k
sinhkH
2
.
The system (4.2) can be rewritten in the form of (2.1), with n = 2 and h = (h
1
, h
2
), by
putting
N(h) =
1
6
(
1

4
h
3
1
,
2

2
h
3
2
)
28
and
D =
_

4
0
0
2
_
,
and dening the symbol A(k) of L by
A(k) =
_

4
(c
1
+
1
m
1
(k) +
2
m
2
(k))
2

4
n(k)

4
n(k)
2
(c
2
+
4
m
2
(k) +
3
m
3
(k))
_
.
In Theorem 2.3 of [3] it is shown that (4.2) is globally well-posed in Y
r
for any r 3/2.
As in the case of (4.1), the functional N is homogeneous of degree 3, so we take p = 1.
The eigenvalues of D +A(k) are given by (cf. [3])

1
(k, ) =
T(k)
2

_
T(k)
2
4d(k)

2
(k, ) =
T(k)
2
+
_
T(k)
2
4d(k),
where T(k) is the trace of D+A(k) and d(k) is the determinant of D+A(k). It follows
easily from the properties of m
i
(k) and n(k) that (A3) (and hence also (A1) and (A2)) is
satised with s = 1/2, and that
0
is the larger of the two roots of the equation
( +c
1
)( +c
2
) =

2

4
H
2
2
;
i.e.

0
=
1
2
_
(c
1
+c
2
) +

(c
1
c
2
)
2
+
4
2

4
H
2
2
_
. (4.3)
Moreover,
0
D + A(k) has the eigenvalue 0 at k = 0, so that (2.4) holds with s
0
= 1, by
the rst remark following Theorem 2.2. Also, N(h) = N(h), so that the second remark
following Theorem 2.2 applies. It follows that all the assumptions underlying Theorem 2.2
are satised, and we obtain the following result.
Theorem 4.2. Let E and Q be the invariant functionals associated with (4.2), as dened
in Section 2. Then for each q > 0, the problem of minimizing E subject to the constraint
Q = q has a nonempty solution set G
q
, and for each g G
q
, there exists c >
0
such
that g(x ct) is a solution of (4.2). Moreover, the set G
q
is stable in the sense that for
every > 0, there exists > 0 with the following property: if h
0
is any function in Y
3/2
satisfying
|h
0
g|
1/2
<
29
for some g G
q
, then there exists a global solution h(x, t) of (4.2) with h(x, 0) = h
0
and
a map t g(t) from [0, ) to G
q
such that
|h(, t) g(t)|
1/2
<
for all t 0.
5. Symmetry of Liu-Kubota-Ko solitary waves.
We begin this section with a lemma that establishes a correspondence between solitary-
wave solutions of (4.2) and solutions of a certain nonlinear boundary-value problem for the
Laplacian, posed on the three innite strips S
1
, S
2
, S
3
dened as subsets of R
2
by
S
1
= R [0, H
1
],
S
2
= R [H
2
, 0],
S
3
= R [(H
2
+H
3
), H
2
].
Lemma 5.1. Let = (
1
,
2
) X
2
be such that (x ct) solves (4.2) for some c >
0
,
where
0
is as dened in (4.3). Then there exist functions u
i
C

(S
i
), i 1, 2, 3, such
that
(i) for i = 1, 2, 3, u
i
= 0 on S
i
,
(ii) for i = 1, 2, 3, u
i
(x, y) 0 uniformly in y as [x[ ,
(iii) u
1
= 0 for y = H
1
,
(iv) u
3
= 0 for y = (H
2
+H
3
),
(v) u
1
= u
2
=
1
for y = 0,
(vi) u
2
= u
3
=
2
for y = H
2
,
(vii)
_
(c +c
1
) +

1
H
1
+

2
H
2
_

1
+

1
2

2
1
+
1
u
1y

2
u
2y
= 0 for y = 0,
(viii)
_
(c +c
2
) +

3
H
3
+

4
H
2
_

2
+

2
2

2
2
+
4
u
2y

3
u
3y
= 0 for y = H
2
.
Proof. As shown in Lemmas 4.2 and 4.3 of [3], if X
2
is any solitary-wave solution of
(4.2) with wavespeed c >
0
, then must in fact be in Y

. Therefore, if we dene u
1
on
30
S
1
, u
2
on S
2
, and u
3
on S
3
by the formulas
u
1
(x, y) =
1
2
_

e
ikx
_
sinhk(H
1
y)
sinhkH
1
_

1
(k) dk,
u
2
(x, y) =
1
2
_

e
ikx
__
sinhk(H
2
+y)
sinhkH
2
_

1
(k)
_
sinh ky
sinhkH
2
_

2
(k)
_
dk,
u
3
(x, y) =
1
2
_

e
ikx
_
sinhk(H
2
+H
3
+y)
sinhk(H
2
+H
3
)
_

2
(k) dk,
it follows from standard arguments (see the proof of Lemma 2 in [5]) that u
i
C

(S
i
)
and tends to 0 uniformly in y as [x[ , and that the partial derivatives of u
i
on S
i
may
be computed by dierentiating under the integral. In particular, dierentiation under the
integral shows that (i) holds, and also that
u
1y

y=0
= M
1

1

1
H
1

1
,
u
2y

y=0
= S
2
+M
2

1
+
1
H
2

1
,
u
2y

y=H
2
= S
1
M
2

2

1
H
2

2
,
u
3y

y=H
2
= M
3

2
+
1
H
3

2
.
Substitution of these expressions in the solitary-wave equation for yields (vii) and (viii).
Finally, (iii)(vi) are obvious from the denitions of the functions u
i
.
Remark. The converse of Lemma 5.1 holds, in the following sense. For arbitrary Y
1
,
there are unique harmonic functions u
i
dened on the interior of S
i
such that [u
i
(, y)[
L
2
is uniformly bounded in y, and
lim
y0
u
1
= lim
y0
u
2
=
1
,
lim
yH
2
u
2
= lim
yH
2
u
3
=
2
,
where the limits are taken in the L
2
sense. The derivatives u
iy
are also well-dened as L
2
traces on y = 0 and y = H
2
. If (vii) and (viii) hold, then (x ct) is a solution of
(4.2).
We will work below not with the functions u
i
themselves, but instead with functions u
i
which we now proceed to dene. The assumption c >
0
implies that c +c
1
and c +c
2
are
31
positive and satisfy
(c +c
1
)(c +c
2
) >

2

4
H
2
2
.
Therefore it is possible to nd a number
2
such that 1 +
2
H
2
is positive and satises

4
(c +c
2
)H
2
< 1 +
2
H
2
<
(c +c
1
)H
2

2
.
Once
2
has been chosen, we can choose
1
and
3
such that 1+
1
H
1
and 1+
2
H
2
+
3
H
3
are positive and satisfy
(1 +
1
H
1
) <

2
H
1

1
H
2
_
(c +c
1
)H
2

2
(1 +
2
H
2
)
_
(5.1)
and
(1 +
2
H
2
+
3
H
3
) <
(c +c
2
)H
3

3
_
(1 +
2
H
2
)

4
(c +c
2
)H
2
_
. (5.2)
Dene
g
1
(y) = 1 +
1
y for 0 y H
1
,
g
2
(y) = 1
2
y, for H
2
y 0,
g
3
(y) = 1 +
2
H
2

3
(y +H
2
) for (H
2
+H
3
) y H
2
.
Notice that g
1
(0) = g
2
(0) and g
2
(H
2
) = g
3
(H
2
), and that each function g
i
(y) takes
only positive values on its domain. Hence we may dene functions u
i
on S
i
for i = 1, 2, 3
by
u
i
(x, y) = g
i
(y) u
i
(x, y).
Properties (ii), (iii), and (iv) of Lemma 5.1 still hold with u
i
replaced by u
i
, and it is
still true, as in (v) and (vi), that u
1
= u
2
for y = 0 and u
2
= u
3
for y = H
2
. Also,
although the functions u
i
are no longer harmonic, they do satisfy the elliptic equation
u
i
+
2g

i
(y)
g
i
(y)
u
iy
= 0 (5.3)
on S
i
. Finally, we see from equations (vii) and (viii) that
Q
1
u
2
+

1
2
u
2
2
+
1
u
1y

2
u
2y
= 0 for y = 0 (5.4)
32
and
Q
2
u
2
+

2
2
u
2
2
+
4
(1 +
2
H
2
) u
2y

3
(1 +
2
H
2
) u
3y
= 0 for y = H
2
, (5.5)
where
Q
1
= (c +c
1
) +

1
H
1
(1 +
1
H
1
) +

2
H
2
(1 +
2
H
2
),
Q
2
= (c +c
2
)(1 +
2
H
2
) +

4
H
2
+

3
H
3
(1 +
2
H
2
+
3
H
3
).
From (5.1) and (5.2), we see that both Q
1
and Q
2
are negative. Notice that such was not
necessarily the case for the coecients of
1
and
2
in (vii) and (viii) of Lemma 5.1. This
is the reason for working with u
i
instead of u
i
.
In what follows, we make repeated use of the fact that solutions of (5.3) satisfy a
maximum principle: if is a bounded connected domain in R
2
, and u C
2
() C
0
(

)
satises (5.3) on , then u must achieve its maximum and minimum values over

on
the boundary of . Further, if either of these values is attained at any point within ,
then u is constant on (see Theorem 3.5 of [22]). On an unbounded domain such as S
i
,
similar assertions can be made in the presence of additional assumptions on the behavior
of u at innity. For example, suppose u satises (5.3) on S
i
and u 0, uniformly in y, as
[x[ . By applying the maximum principle on sets S
i
R x R as R , we
can deduce that if u takes a negative value anywhere on S
i
, then the minimum value of u
over S
i
must be attained at some point on the boundary of S
i
.
We will also use the following renements of the maximum principle, which are valid
on any domain R
2
, bounded or unbounded. The Hopf boundary lemma implies that
if u satises (5.3) on and attains its minimum value over

at a point (x
0
, y
0
) on the
boundary of , and there exists a ball in whose boundary contains (x
0
, y
0
), then the
normal derivative of u at (x
0
, y
0
) is zero only if u is constant on

(see Lemma 3.4 of
[22]). There is also a Hopf corner-point lemma [21], which has the following implication
for (5.3). Suppose u C
2
() C
0
(

) satises (5.3) on , where is the semi-innite

strip (, x
0
) (y
0
, y
1
), and u is non-negative on and tends to 0, uniformly in y as
x . Let P be a corner point of ; i.e., P = (x
0
, y
0
) or P = (x
0
, y
1
). If u = 0 at P,
and the (one-sided) derivatives u
x
, u
y
, u
xx
, u
xy
, and u
yy
exist and are all equal to 0 at P,
then u is identically zero on .
33
Lemma 5.2. Suppose
1
and
2
are positive. Then u
i
0 on S
i
for i = 1, 2, 3.
Proof. Suppose, to the contrary, that for some i, u
i
takes a negative value at some point
of S
i
. We will show that then either (5.4) or (5.5) must fail to hold. Since u
i
satises (5.3)
on S
i
and tends to 0 uniformly in y as [x[ , it follows from the maximum principle
that the minimum value of u
i
on S
i
must be attained at some point on the boundary of
S
i
. Furthermore, since u
1
= 0 for y = H
1
and u
3
= 0 for y = (H
2
+ H
3
), this negative
minimum can only be attained on the boundary of S
2
, where u
i
= u
2
. Hence u
2
must take
a negative minimum value at some point (x
0
, y
0
) on the boundary of S
2
.
There are now two possibilities: either y
0
= 0 or y
0
= H
2
. In the rst case, (x
0
, 0) is a
minimum value both for u
1
on S
1
and for u
2
on S
2
, so we must have that u
1y
(x
0
, 0) 0 and
u
2y
(x
0
, 0) 0. On the other hand, Q
1
< 0 and u
2
(x
0
, 0) < 0. Combining these facts, we
conclude that the left-hand side of equation (5.4) is strictly positive, so (5.4) is contradicted.
An exactly similar argument shows that if y
0
= H
2
, then (5.5) is contradicted. Thus the
proof is complete.
Remark. From Theorem 5.4(ii) below it follows that if
1
and
2
are positive and the
functions u
i
are not all identically zero, then
1
(x) and
2
(x) are in fact strictly positive
functions of x (and hence the u
i
are strictly positive at all points in their domains except
where y = H
1
and y = (H
2
+ H
3
)). This fact is also a direct consequence of the proof
of Lemma 5.1, since if
1
or
2
vanishes at some point, then applying the Hopf boundary
lemma to the appropriate u
i
at that point yields a contradiction to (5.4) or (5.5). Yet
another proof of the positivity of
1
and
2
, which does not use Lemma 5.1 at all, is the
following. In Lemmas 4.2 and 4.3 of [3] it is shown that for c >
0
, the operator L + cD
has an inverse (L + cD)
1
, dened on all of X
2
, whose entries are convolution operators
with positive kernels. Now, if
1
and
2
are positive and is not identically zero, then
the entries of N() are non-negative and not identically zero. Hence the function (L +
cD)
1
(N()) is everywhere positive on R. The desired result then follows immediately
upon rewriting the solitary-wave equation (2.2) in the form = (L +cD)
1
(N()).
Now for each R, dene w
i
(x, y, ) for (x, y) S
i
, i = 1, 2, 3, by
w
i
(x, y, ) = u
i
(2 x, y) u
i
(x, y).
34
We will examine the behavior of w
i
on the set

i
() = (x, y) S
i
: x .
Lemma 5.3. Suppose
1
and
2
are positive. Then there exists
0
R such that
(i) for all R, if w
2
(x, y, ) attains a minimum value over
2
() at some point (x
0
, y
0
)
in
2
(), then either x
0
>
0
or w
2
(x
0
, y
0
, ) 0.
(ii) for all
0
and all i 1, 2, 3, we have w
i
(x, y, ) 0 for all (x, y)
i
().
Proof. Let functions B
1
: R
3
R and B
2
: R
3
R be dened by
B
1
(p, q, r) = Q
1
p +

1
2
p
2
+
1
q
2
r
and
B
2
(p, q, r) = Q
2
p +

2
2
p
2
+
4
(1 +
2
H
2
)q
3
(1 +
2
H
2
)r,
so that (5.4) and (5.5) take the form
B
1
( u
2
(x, 0), u
1y
(x, 0), u
2y
(x, 0)) = 0 for all x R (5.6)
and
B
2
( u
2
(x, H
2
), u
2y
(x, H
2
), u
3y
(x, H
2
)) = 0 for all x R. (5.7)
Since Q
1
and Q
2
are negative, and u
2
0 uniformly in y as x , we can nd
0
such
that if x
0
, then
B
1
p
= Q
1
+
1
p < 0 at p = u
2
(x, 0) (5.8)
and
B
2
p
= Q
2
+
2
p < 0 at p = u
2
(x, H
2
). (5.9)
We now prove (i) by contradiction. Suppose x
0

0
and w
2
(x
0
, y
0
, ) < 0. Since the
functions w
i
satisfy the same equation (5.3) on S
i
as do the functions u
i
, and w
i
0 as
x , we conclude from the maximum principle that (x
0
, y
0
) must lie on the boundary
of
2
(). Moreover, since w
2
(, y, ) = 0 for all y, (x
0
, y
0
) must lie on the horizontal part
of the boundary of
2
(), so that either y
0
= 0 or y
0
= H
2
.
35
Consider rst the case y
0
= 0. Since w
2
(x, y, ) attains its minimum value on
2
() at
(x
0
, 0), we have w
2y
(x
0
, 0, ) 0, and hence
u
2y
(2 x
0
, 0) u
2y
(x
0
, 0). (5.10)
Also, since w
1
= w
2
for y = 0, the maximum principle implies that the minimum value of
w
1
(x, y, ) on
1
() is also attained at (x
0
, 0). Therefore w
1y
(x
0
, 0, ) 0, and so
u
1y
(2 x
0
, 0) u
1y
(x
0
, 0). (5.11)
On the other hand, since w
2
(x
0
, 0, ) < 0, we have
u
2
(2 x
0
, 0) < u
2
(x
0
, 0). (5.12)
Since
1
> 0, it follows from (5.8) and (5.12) that
B
1
p
= Q
1
+
1
p < 0 for all p [ u
2
(2 x
0
, 0), u
2
(x, 0)]. (5.13)
Finally, combining (5.10), (5.11), (5.12) and (5.13), and recalling that
1
and
2
are posi-
tive, we obtain that
B
1
( u
2
(2 x
0
, 0), u
1y
(2 x
0
, 0), u
2y
(2 x
0
, 0)) > B
1
( u
2
(x
0
, 0), u
1y
(x
0
, 0), u
2y
(x
0
, 0)),
contradicting (5.6).
We have shown that in the case y
0
= 0, we obtain a contradiction to (5.6). On the other
hand, if y
0
= H
2
, an exactly similar argument leads to a contradiction of (5.7). Thus
the proof of (i) is complete.
To prove (ii), suppose to the contrary that for some i and some
0
, w
i
(x, y, )
takes a negative value on
i
(). Then since w
i
0 as x , the maximum principle
implies that w
i
(x, y, ) attains its minimum over
i
() at some point on the boundary
shared by
i
() and
2
(). Hence w
2
takes a negative value on
2
(), and so attains a
negative minimum value over
2
() at some point (x
0
, y
0
) in
2
(). Since x
0

0
,
this contradicts (i).
We can now state and prove the main theorem of this section.
36
Theorem 5.4. Suppose that
1
and
2
are positive, and the u
i
are not identically zero.
Then there exists R such that
(i) for all i 1, 2, 3 and all (x, y) S
i
,
u
i
(x, y) = u
i
(2 x, y).
(ii) for all i 1, 2, 3 and all (x, y) S
i
such that x < and (H
2
+H
3
) < y < H
1
,
u
i
x
(x, y) > 0.
Proof. Dene the number by
= sup : if , then w
i
(x, y, ) 0 for all (x, y)
i
() and all i 1, 2, 3 .
Lemma 5.4(ii) shows that > , and it follows easily from Lemma 5.1(ii) and Lemma
5.2 that < .
To prove (i), it suces to show that for each i = 1, 2, 3 we have w
i
(x, y, ) = 0 for all
(x, y)
i
( ).
By the denition of , we can nd a sequence
k
such that
k
and for each k,
there exists i 1, 2, 3 for which w
i
(x, y,
k
) takes a negative minimum on
i
(
k
). As
noted in the proof of Lemma 5.3, it follows that w
2
(x, y,
k
) must take a negative minimum
value on
2
(
k
), and this value must be achieved at a point (x
k
, y
k
) where either y
k
= 0
or y
k
= H
2
. By passing to a subsequence, we may assume that either y
k
= 0 for all k, or
y
k
= H
2
for all k. We will consider the former of these two cases, the proof in the latter
case being exactly similar.
From Lemma 5.3(i) it follows that x
k
>
0
for all k. Therefore the sequence x
k

is bounded, so by again passing to a subsequence we may assume that x

k
converges to
some number x . Since w
2
(x
k
, 0,
k
) < 0 for all k, then w
2
( x, 0, ) 0. Hence also
w
1
( x, 0, ) 0. On the other hand, from the denition of it follows that w
1
(x, y, ) 0
on
1
( ) and w
2
(x, y, ) 0 on
2
( ). Therefore
w
1
( x, 0, ) = w
2
( x, 0, ) = 0, (5.14)
37
and it follows that
w
1y
( x, 0, ) 0 (5.15)
and
w
2y
( x, 0, ) 0. (5.16)
We now consider separately the cases when x < and when x = . Suppose rst that
x < . Substitute into (5.4) the values x = 2 x and x = x, and subtract the two
resulting equations, using (5.14). There appears the identity

1
w
1y
( x, 0, ) =
2
w
2y
( x, 0, ),
which together with (5.15) and (5.16) yields
w
1y
( x, 0, ) = w
2y
( x, 0, ) = 0.
Hence, applying the Hopf boundary lemma to w
2
(x, y, ) at the point ( x, 0), we obtain
that w
2
(x, y, ) is identically zero on
2
( ). It then follows from the maximum principle
that w
1
(x, y, ) is identically zero on
1
( ) and w
3
(x, y, ) is identically zero on
3
( ).
Thus (i) has been proved in case x < , and we may assume henceforth that x = .
We proceed to investigate the derivatives of w
2
(x, y, ) at the point (x, y) = ( , 0).
Notice rst that since w
2
(x, y,
k
) attains a minimum over
2
(
k
) at (x
k
, 0), and x
k
<
k
,
we must have w
2x
(x
k
, 0,
k
) = 0 for all k. Taking the limit as k then gives
w
2x
( , 0, ) = 0. (5.17)
Since, for any and y,
w
ix
(, y, ) = 2 u
ix
(, y), (5.18)
then (5.17) implies
u
2x
( , 0) = 0. (5.19)
Also, clearly w
2
( , y, ) = 0 for H
2
y 0, so
w
2y
( , 0, ) = w
2yy
( , 0, ) = 0. (5.20)
38
Since w satises (5.3), then (5.20) implies
w
2xx
( , 0, ) = 0. (5.21)
Finally we consider the mixed derivative w
2xy
( , 0, ). Observe that since the minimum
value of w
1
(x, y, ) on
1
( ) is attained at every point ( , y) in
1
( ), then w
1x
( , y, ) 0
for 0 y H
1
. Similarly, we have w
2x
( , y, ) 0 for H
2
y 0. Taking (5.17) into
account, we conclude that w
1xy
( , 0, ) 0 and w
2xy
( , 0, ) 0, so that from (5.18) we
obtain
u
1xy
( , 0) 0 (5.22)
and
u
2xy
( , 0) 0. (5.23)
On the other hand, dierentiating (5.4) with respect to x and evaluating at x = using
(5.17), we obtain

1
u
1xy
( , 0) =
2
u
2xy
( , 0),
which, together with (5.22) and (5.23), implies
u
1xy
( , 0) = u
2xy
( , 0) = 0.
In particular, it follows that
w
2xy
( , 0, ) = 0. (5.24)
We have now shown (in (5.14), (5.17), (5.20), (5.21), and (5.24)) that, at the point
(x, y) = ( , 0), w
2
(x, y, ) and all its partial derivatives of rst and second order are equal
to zero. It therefore follows from the Hopf corner-point lemma that w
2
(x, y, ) must be
zero for all (x, y) in
2
( ). As above, this is enough to conclude that (i) holds. Hence (i)
has now been proved in all cases.
To prove (ii), we rst note that if the u
i
are not identically zero, then there does not
exist any < such that w
2
(x, y, ) is identically zero on
2
(). For if there were such a
, it would follow that the u
i
are symmetric about as well as , and hence that the u
i
are periodic of period , contradicting the fact that u
i
0 as [x[ .
39
Now observe that, if it were the case that w
2x
(, 0, ) = 0 for some < , then the same
chain as reasoning as above, starting with (5.17) and concluding with (5.24), would show
that w
2
(x, y, ) and all its partial derivatives up to second order are zero at (x, y) = (, 0).
The Hopf corner-point lemma would then imply that w
2
(x, y, ) is identically zero on

2
(), contradicting the result of the preceding paragraph. Therefore w
2x
(, 0, ) ,= 0.
However, since w
i
(x, y, ) 0 on
i
() and w
i
(, y, ) = 0, we must have
w
ix
(, y, ) 0 for all (, y)
i
(). (5.25)
Hence w
2x
(, 0, ) < 0, so by (5.18) we have u
2x
(, 0) > 0. It follows that u
1x
(, 0) > 0
also. A similar argument shows that u
2x
(, H
2
) = u
3x
(, H
2
) > 0.
It remains to show that u
ix
(, y) > 0 if (H
2
+H
3
) < y < H
1
and y is neither H
2
nor
0. By (5.25), it suces to show that w
ix
(, y, ) cannot be zero for such y. But if indeed
w
ix
(, y, ) = 0 for some i, then since (, y) is on the interior of the vertical boundary
of
i
(), the Hopf boundary lemma implies that w
i
(x, y, ) is identically zero on
i
().
We know from above that i cannot equal 2, so either i = 1 or i = 3. But in either case,
the fact that w
i
(x, y, ) is identically zero on
i
() implies that w
2x
(x, y, ) = 0 at one of
the corner points of
2
(), and we are back to the situation of the preceding paragraph.
Hence, in any case, we obtain a contradiction, and the proof of (ii) is complete.
Remark. If (x ct) solves (4.2), then (x ct) solves (4.2) with
1
and
2
replaced by

1
and
2
. Hence it follows from Theorem 5.4 that if
1
and
2
are negative, then the
conclusions of the theorem hold for . We do not yet know an analogue of Theorem 5.4
in the case when
1
and
2
have dierent signs.
It is possible to extend Theorem 5.4 to cases in which the u
i
do not tend to zero in both
horizontal directions.
Theorem 5.5. Suppose that
1
and
2
are positive. Suppose also that functions u
i

C
2
(S
i
) are given which satisfy all the conditions of Lemma 5.1, except that (ii) is replaced
by the requirements that, for some > 0 and all i 1, 2, 3, we have
lim
x
u
i
= 0 and lim
x+
u
i
= ,
40
both limits being uniform in y. Then for all i 1, 2, 3 and all (x, y) S
i
such that
(H
2
+H
3
) < y < H
1
, we have
u
i
x
(x, y) > 0.
Proof. Dene u
i
and w
i
as before. Under the given assumptions on u
i
, the proof of Lemma
5.2 still goes through, showing that u
i
0 on S
i
for i = 1, 2, 3. Also, although it is now
no longer necessarily the case that w
i
0 as x , it is still true that
liminf
x
w
i
(x, y, ) 0,
and it may be easily checked that this is sucient for the proof of Lemma 5.3 to be carried
out as before. Thus Lemma 5.3 still holds, and in particular we may dene as before
with the assurance that > .
Now if < , then the proof of Theorem 5.4(i) shows that the u
i
are symmetric about
, which contradicts our assumptions about the behavior of u
i
as x . Therefore
we must have = . The desired result is now obtained by noticing that the proof of
Theorem 5.4(ii) goes through unchanged in the present situation.
Remark. The functions u
i
described in Theorem 5.5 do not arise from L
2
solitary-wave
solutions of (4.2): indeed, since
1
(x) = u
1
(x, 0) and
2
(x) = u
2
(x, H
2
) are not in L
2
,
the operator L will not in general be well-dened at = (
1
,
2
). On the other hand, if
one were to derive equations modeling bore-like waves at the interfaces of a three-layered
uid by a procedure analogous to the one used to derive (4.2) for localized waves, then
would represent a valid solution to such a system. Thus Theorem 5.5 can be interpreted
as a result for a system of equations modelling internal bores.
Finally we note that the above arguments also yield a symmetry result for solitary-wave
solutions of a scalar equation derived by Kubota, Ko and Dobbs [24] as a model for long
waves in a stratied uid at the interface between two layers of constant density, one layer
having (non-dimensionalized) depth equal to H
1
and the other layer having depth H
2
.
After a suitable choice of variables, the Kubota-Ko-Dobbs equation may be put in the
form
h
t
+hh
x

1
(M
1
h)
x

2
(M
2
h)
x
= 0, (5.26)
41
where h : RR R and
1
,
2
are positive real numbers. The operators M
1
and M
2
are
dened as in Section 4b, and in place of (4.3) we now have
0
= 0. In case H
1
= H
2
, (5.26)
is known as the Intermediate Long Wave (ILW) equation, and has been extensively studied
in the literature devoted to completely integrable equations (cf. [1]). The ILW equation
has, for each c > 0, a solitary-wave solution (xct) given by an explicit formula in terms
of exponential functions; in particular this formula shows that is symmetric about the
point where it attains its maximum value and is strictly decreasing away from that point.
Further, it is known [5] that is, up to translation, the unique solitary wave solution of
ILW with wavespeed c. In case H
1
,= H
2
, it is known (see Theorem 3.1 of [3]) that for each
c > 0, (5.26) has at least one solitary-wave solution (x ct) which is symmetric about
its maximum, but it remains an open question whether is unique up to translation. The
following result is therefore of interest.
Theorem 5.6. Let L
2
(R) be such that (x ct) solves (5.26) for some c > 0. Then
there exists R such that (2 x) = (x) for all x R and

(x) > 0 for all x < .

To prove Theorem 5.6, one rst observes that solitary-wave solutions of (5.26) are asso-
ciated with the same kind of elliptic boundary-value problem as specied in Lemma 5.1,
except that now only two strips are involved. More precisely, let S
1
and S
2
be the innite
strips in R
2
dened by
S
1
= R [0, H
1
],
S
2
= R [H
2
, 0].
Then for as in the statement of Theorem 5.6, one nds that there exist functions u
1

C

(S
1
) and u
2
C

(S
2
) such that
(i) u
1
= 0 on S
1
, and u
2
= 0 on S
2
,
(ii) u
1
and u
2
tend to 0 uniformly in y as [x[ ,
(iii) u
1
= 0 for y = H
1
,
(iv) u
2
= 0 for y = H
2
,
(v) u
1
= u
2
= for y = 0, and
(vi)
_
c +

1
H
1
+

2
H
2
_
+
1
2

2
+
1
u
1y

2
u
2y
= 0 for y = 0.
42
Now choose
1
and
2
such that 1 +
1
H
1
and 1 +
2
H
2
are positive numbers, and
Q = c +

1
H
1
(1 +
1
H
1
) +

2
H
2
(1 +
2
H
2
) < 0.
Dening g
1
, g
2
and u
1
, u
2
by the same formulas as given above prior to Lemma 5.2, we
nd that (vi) implies
Q u
2
+
1
2
u
2
2
+
1
u
1y

2
u
2y
= 0 for y = 0.
From here the proof of Theorem 5.6 proceeds exactly like the proof of Theorem 5.4, and
we can safely omit the details.
6. Acknowledgement.
The authors wish to thank Hongqiu Chen for an important conversation.
References
[1] M. Ablowitz and H. Segur, Solitons and the inverse scattering transform, SIAM, Philadelphia, PA,
1981.
[2] J. Albert, Concentration compactness and the stability of solitary-wave solutions to nonlocal equa-
tions, Applied analysis (Baton Rouge, LA, 1996), Contemporary Mathematics vol. 221, American
Mathematical Society, Providence, RI, 1999, pp. 129.
[3] J. Albert, J. Bona, and J.-C. Saut, Model equations for waves in stratied uids, Proc. Roy. Soc.
London Ser. A 453 (1997), 12331260.
[4] J. Albert and F. Linares, Stability of solitary-wave solutions to long-wave equations with general
dispersion, Matematica Contemporanea 15 (1998), 119.
[5] J. Albert and J. Toland, On the exact solutions of the intermediate long-wave equation, Dierential
Integral Equations 7 (1994), 601612.
[6] J. Angulo Pava, Existence and stability of solitary wave solutions of the Benjamin equation, J. Dif-
ferential Equations 152 (1999), 136159.
[7] D. Benney, Long non-linear waves in uid ows, J. Math. and Physics 45 (1966), 5263.
[8] J. Bergh and J. Lofstrom, Interpolation spaces: an introduction, Springer-Verlag, New York, 1976.
[9] J. Bona, On solitary waves and their role in the evolution of long waves, Applications of Nonlinear
Analysis in the Physical Sciences (H. Amann, N. Bazley, and K. Kirchg assner, eds.), Pitman, New
York, 1981, pp. 183205.
[10] J. Bona and H. Chen, Solitary waves in stratied uid systems, to appear.
[11] J. Bona, G. Ponce, J.-C. Saut, and M. Tom, A model system for strong interaction between internal
solitary waves, Comm. Math. Phys. 143 (1992), 287313.
[12] J. Boussinesq, Theorie des ondes et des remous qui se propagent le long dun canal rectangulaire
horizontal, en communiquant au liquide contenu dans ce canal des vitesses sensiblement pareilles de
la surface au fond, J. Math. Pures Appl. (2) 17 (1872), 55108.
[13] T. Cazenave and P.-L. Lions, Orbital stability of standing waves for some nonlinear Schrodinger
equations, Comm. Math. Phys. 85 (1982), 549561.
[14] H. Chen and J. Bona, Existence and asymptotic properties of solitary-wave solutions of Benjamin-type
equations, Adv. Dierential Equations 3 (1998), 5184.
[15] R. Coifman and Y. Meyer, Au del`a des operateurs pseudo-dierentiels, Asterisque no. 57, Societe
Mathematique de France, Paris, 1978.
43
[16] W. Craig and P. Sternberg, Symmetry of solitary waves, Comm. P.D.E. 13 (1988), 603633.
[17] W. Craig and P. Sternberg, Symmetry of free-surface ows, Arch. Rational Mech. Anal. 118 (1992),
136.
[18] A. de Bouard and J.-C. Saut, Remarks on the stability of generalized KP solitary waves, Mathematical
problems in the theory of water waves (Luminy, 1995), Contemporary Mathematics vol. 200, American
Mathematical Society, Providence, RI, 1996, pp. 7584.
[19] A. de Bouard and J.-C. Saut, Solitary waves of generalized Kadomtsev-Petviashvili equations, Ann.
Inst. H. Poincare Anal. Non Lineaire 14 (1997), 211236.
[20] J. Gear and R. Grimshaw, Weak and strong interactions between internal solitary waves, Stud. Appl.
Math 70 (1984), 235258.
[21] B. Gidas, W.-M. Ni, and L. Nirenberg, Symmetry and related properties via the maximum principle,
Comm. Math. Phys. 68 (1979), 209243.
[22] D. Gilbarg and N. Trudinger, Elliptic partial dierential equations of second order, Springer-Verlag,
Berlin-New York, 1983.
[23] T. Kato, Perturbation theory for linear operators, Springer-Verlag, Berlin-New York, 1976.
[24] T. Kubota, D. Ko, and L. Dobbs, Weakly nonlinear interval gravity waves in stratied uids of nite
depth, J. Hydrodynamics 12 (1978), 157165.
[25] C. Li, Monotonicity and symmetry of solutions of fully nonlinear elliptic equations on unbounded
domains, Comm. P.D.E 16 (1991), 585615.
[26] P. Lions, The concentration compactness principle in the calculus of variations. The locally compact
case, part 1, Ann. Inst. H. Poincare 1 (1984), 109145.
[27] A. Liu, T. Kubota, and D. Ko, Resonant transfer of energy between nonlinear waves in neighboring
pycnoclines, Stud. Appl. Math. 63 (1980), 2545.
[28] D. Luenberger, Optimization by vector space methods, Wiley and Sons, New York, 1969.
[29] L. Maia, Symmetry of internal waves, Nonlinear Analysis TMA 28 (1997), 87102.
[30] R. Pego and M. Weinstein, Asymptotic stability of solitary waves, Comm. Math. Phys. 164 (1994),
305349.
[31] M. Weinstein, Existence and dynamic stability of solitary wave solutions of equations arising in long
wave propagation, Commun. Partial Di. Equations 12 (1987), 11331173.
Department of Mathematics, University of Oklahoma, Norman, OK 73019, USA
E-mail address: [email protected]
IMPA, Estrada Dona Castorina 110, 22460-320 Rio de Janeiro RJ, Brasil
E-mail address: [email protected]