A numerical scheme based on multipeakons for

conservative solutions of the CamassaHolm

equation
Helge Holden
1,2
and Xavier Raynaud
1
1
Department of Mathematical Sciences, Norwegian University of Science and
Technology, NO7491 Trondheim, Norway;

raynaud@math.ntnu.no
2
Centre of Mathematics for Applications, Department of Mathematics, University
of Oslo, P.O. Box 1053, Blindern, NO0316 Oslo, Norway; holden@math.ntnu.no
Summary. We present a convergent numerical scheme based on multipeakons to
compute conservative solutions of the CamassaHolm equation.
1 Introduction
The CamassaHolm equation [3, 4] reads
u
t
u
xxt
+ 3uu
x
2u
x
u
xx
uu
xxx
= 0. (1)
It has a bi-Hamiltonian structure, is completely integrable, and has innitely
many conserved quantities. In this paper we study the Cauchy problem
u|
t=0
= u H
1
(R) for (1). A highly interesting property of the equation
is that for a wide class of initial data the solution experiences wave break-
ing in nite time in the sense that the solution u remains bounded pointwise
while the spatial derivative u
x
becomes unbounded pointwise. However, the
H
1
norm of u remains nite, see, e.g., [7, 8].
The extension of the solution beyond wave breaking is non-trivial and can
be illustrated by studying multipeakons that are solutions of the form
u(t, x) =
n

i=1
p
i
(t)e
|xqi(t)|
, (2)
where the (p
i
(t), q
i
(t)) satisfy the explicit system of ordinary dierential equa-
tions
q
i
=
n

j=1
p
j
e
|qiqj|
, p
i
=
n

j=1
p
i
p
j
sgn(q
i
q
j
)e
|qiqj|
. (3)

The research is supported by the Research Council of Norway.

2 Holden and Raynaud
Observe that the solution (2) is not smooth even with continuous functions
(p
i
(t), q
i
(t)).
Peakons interact in a way similar to that of solitons of the Korteweg
de Vries equation, and wave breaking may appear when at least two of the
q
i
s coincide. If all the p
i
(0) have the same sign, the peakons move in the
same direction and we have a global solution and no wave breaking. Higher
peakons move faster than the smaller ones, and when a higher peakon over-
takes a smaller, there is an exchange of mass, but no wave breaking takes
place. Furthermore, the q
i
(t) remain distinct. However, if some of p
i
(0) have
opposite sign, wave breaking may incur. For simplicity, consider the case with
n = 2 and one peakon p
1
(0) > 0 (moving to the right) and one antipeakon
p
2
(0) < 0 (moving to the left). In the symmetric case (p
1
(0) = p
2
(0) and
q
1
(0) = q
2
(0) < 0) the solution will vanish pointwise at the collision time
t

when q
1
(t

) = q
2
(t

), that is, u(t

, x) = 0 for all x R. Clearly, at least

two scenarios are possible; one is to let u(t, x) vanish identically for t > t

,
and the other possibility is to let the peakon and antipeakon pass through
each other in a way that is consistent with the CamassaHolm equation. In
the rst case the energy
_
(u
2
+ u
2
x
) dx decreases to zero at t

, while in the
second case, the energy remains constant except at t

. The rst solution is

denoted a dissipative solution, while the second one is called conservative.
Other solutions are also possible. Global dissipative solutions of a more gen-
eral class of equations were recently derived by Coclite, Holden, and Karlsen
[6, 5], while global conservative solutions have been derived by Bressan, Fonte,
and Constantin [1, 2], and Holden and Raynaud [9, 10].
Multipeakons are fundamental building blocks for general solutions. In-
deed, if the initial data u is in H
1
and m := u u

is a positive Radon
measure, then it can proved, see [12], that one can construct a sequence of
multipeakons that converges in L

loc
(R; H
1
loc
(R)) to the unique global solution
of the CamassaHolm equation. Here we extend this analysis to avoid the
sign constraint on m for given initial data u. We are thereby able to include
the case where one may have wave breaking, and we select the conservative
solution. This is in part based on [12] and a reformulation of the formulas for
multipeakons [13]. In BressanFonte [2] the approximation of general initial
data by multipeakons is proved in an abstract way, while we in this paper
provide a constructive approach amenable to numerical computations. The
method is illustrated on concrete examples.
A completely dierent numerical approach based on a semi discrete nite
dierence scheme can be found in [11].
2 Description of the method
It turns out to be useful to reformulate the CamassaHolm equation as the
following system
Numerical scheme for the CamassaHolm equation 3
u
t
+uu
x
+P
x
= 0, P P
xx
= u
2
+
1
2
u
2
x
. (4)
In [9] a continuous semigroup of conservative solutions for the Camassa
Holm equation is constructed. The semigroup is dened on the set D of pair
(u, ) where u H
1
(R) and is a Radon measure satisfying
ac
= (u
2
+
u
x
) dx. The conservative solutions are obtained by a change of variable to
Lagrangian coordinates. The new set of coordinates (y, U, H) correspond to
the characteristics, y
t
(t, ) = u(t, y(t, )), the Lagrangian velocity, U(t, ) =
u(t, y(t, )), and the Lagrangian energy distribution, H(t, ) =
_
y(t,)

(u
2
+
u
2
x
) dx, respectively. They satisfy the following system of ordinary dierential
equations in some Banach space E:
y
t
= U, U
t
= Q, H
t
= U
3
2PU, (5)
where
Q(t, ) =
1
4
_
R
sgn( ) exp
_
sgn( )(y() y())
__
U
2
y

+H

_
() d,
P(t, ) =
1
4
_
R
exp
_
sgn( )(y() y())
__
U
2
y

+H

_
d.
The set D is in bijection with a closed subset of E, see [9], and the topology
on D is the topology of E transported to D by this bijection. In [9], we prove
that if u
n
u in H
1
(R) then (u
n
, (u
2
n
+u
2
n,x
) dx) (u, (u
2
+ u
2
x
) dx) and if
(u
n
,
n
) (u, ) in D then u
n
u in L

(R) and
n

.
Multipeakons are particular solutions of the CamassaHolmequation given
by (2)(3) up to collision time, that is, a time t
c
such that q
i
(t
c
) = q
i+1
(t
c
) for
some i. At collision time, the system (3) blows up and we have lim
ttc
p
i
(t) =
lim
ttc
p
i+1
(t) = .
An equivalent denition of a multipeakon with n peaks is given by the
set of pairs (q
i
, u
i
) for i = 1, . . . , n where q
i
represents the position of the ith
peak and u
i
denoted its height. Naturally, we must impose that if q
i
= q
j
for some i and j, then u
i
= u
j
. The multipeakon function u corresponding
to (2) for a given time is then given on the whole real axis as the solution of
the boundary value problem: u u
xx
= 0, u(x
i
) = u
i
, u(x
i+1
) = u
i+1
. We
denote by M
n
the set of all such functions, i.e., functions that correspond
to initial data of multipeakon solutions with n peaks. Furthermore, we set
M =

n=1
M
n
. The two representations are equivalent, and later we explain
how one goes from one to the other. The advantage of the second representa-
tion is that it ts directly into the Lagrangian approach. In [13], we follow this
direction and present a system of ordinary dierential equations for the con-
servative multipeakon solutions of the CamassaHolm equation. Conservative
solutions require that we take into account the energy density (u
2
+ u
2
x
) dx.
We introduce the variables H
i
, for i = 1, . . . , n, which correspond to energy
contained between and q
i
, i.e., H
i
=
_
qi

(u
2
+u
2
x
) dx, and the dierences
H
i
= H
i+1
H
i
. When q
i
= q
i+1
, H
i
is given by
4 Holden and Raynaud
H
i
= H
i+1
H
i
=
_
qi+1
qi
(u
2
+u
2
x
) dx. (6)
Due to the special structure of multipeakons, one can compute H
i
in terms
of q
i
and u
i
. We have
H
i
= 2 u
2
i
tanh(q
i
) + 2u
2
i
coth(q
i
) (7)
where we for convenience have introduced the variables
q
i
=
1
2
(q
i
+q
i+1
), q
i
=
1
2
(q
i+1
q
i
), u
i
=
1
2
(u
i
+u
i+1
), u
i
=
1
2
(u
i+1
u
i
).
In [13], we prove that the structure of multipeakons is preserved by the con-
tinuous semigroup of conservative solutions of the CamassaHolm equation
and the equations for (q
i
, u
i
, H
i
) are given by the following system
dq
i
dt
= u
i
,
du
i
dt
= Q
i
,
dH
i
dt
= u
3
i
2P
i
u
i
, (8)
with P
i
=

n
j=0
P
ij
, Q
i
=

n
j=0

ij
P
ij
, and
P
ij
=
_

_
e
(q1qi)
u
2
1
4
for j = 0,
e

ij
q
i
e

ij
q
j
8 cosh(qj )
_
2H
j
cosh
2
(q
j
)
+8
ij
u
j
u
j
sinh
2
(q
j
) + 4 u
2
j
tanh(q
j
)
for j = 1, . . . , n 1,
e
(qiqn)
u
2
n
4
for j = n.
(9)
The system (8) is well-posed and has global solutions.
Given initial data u H
1
(R), we denote by (u, ) the conservative solution
of the CamassaHolm equation with initial data ( u, ( u
2
+ u
2
x
) dx). We will
prove that there exists a sequence u
n
in M which converges to u in H
1
(R).
We know that ( u
n
, ( u
2
n
+ u
2
n,x
) dx) then converges to ( u, ( u
2
+ u
2
x
) dx) in D.
We compute the multipeakon solution (u
n
,
n
) with initial data ( u
n
, u
2
n
+
u
2
n,x
) dx) by solving the system of ordinary dierential equations (8). Since
the semigroup is continuous, (u
n
,
n
) (u, ) in C([0, T], D) for any T > 0
and therefore u
n
u in L

(R).
3 Approximation of the initial data
In [2], Bressan and Fonte show that M is dense in H
1
(R). The proof is short
and elegant but it is not constructive and therefore not suited to numerical
applications. Here we dene constructive approximation procedure.
For any integer n and any points x
i
, i = 1, . . . , n, such that x
i
x
i+1
,
we denote by P the partition corresponding to those points. Let L
P
be the
span of the partition, i.e., L
P
= x
n
x
1
and
P
the maximum step size, i.e.,
Numerical scheme for the CamassaHolm equation 5

P
= max
i=1,...,n1
(x
i+1
x
i
). Given a partition P = {x
1
, . . . , x
n
}, we dene
by I
P
the operator from H
1
(R) to M dened as follows: For any f H
1
(R),
we set f
i
= f(x
i
) and u = I
P
(f) where u is solution of the boundary problem
u u
xx
= 0, u(x
i
) = f
i
, u(x
i+1
) = f
i+1
(10)
in (x
i
, x
i+1
). The denition (10) extends naturally to the intervals (, x
1
)
and (x
n
, ) by setting x
0
= x
n+1
= and u
0
= u
n+1
= 0. We have the
following theorem.
Theorem 1. For any f H
1
(R), I
P
(f) converges to f in H
1
(R) when
P
tends to zero and L
P
tends to innity.
Proof. Given a partition P, the map I
P
: M H
1
(R) is linear and contin-
uous. Let us prove that I
P
is uniformly continuous, that is, there exists a
constant C such that
I
P
(f)
H
1
(R)
C f
H
1
(R)
(11)
for any f H
1
(R) and any partition P such that
P
1. The H
1
(R) norm
of I
P
(f) can be obtained from (6) and (7):
I
P
(f)
2
H
1
(R)
= f
2
1
+f
2
n
+
n1

i=1
_
2
_
f
i
+f
i+1
2
_
2
tanh
_
x
i+1
x
i
2
_
+ 2
_
f
i+1
f
i
2
_
2
coth
_
x
i+1
x
i
2
__
. (12)
Let = inf
x[xi,xi+1]
f(x)
2
and x
min
be such that f(x
min
)
2
= . Since f is
continuous, x
min
and are well-dened. It follows from the denition of
that
(x
i+1
x
i
) f
2
L
2
(xi,xi+1)
. (13)
We have, for any x [x
i
, x
i+1
],
f(x)
2
= + 2
_
x
xmin
f(t)f

(t) dt + 2 f
2
H
1
(xi,xi+1)

1
x
i+1
x
i
f
2
L
2
(xi,xi+1)
+ 2 f
2
H
1
(xi,xi+1)
,
from (13), which implies f(x)
2
(x
i+1
x
i
) 3 f
2
H
1
(xi,xi+1)
since, by assump-
tion, x
i+1
x
i
1. Hence, f
2
L

(xi,xi+1)
(x
i+1
x
i
) 3 f
2
H
1
(xi,xi+1)
. We
can estimate the rst term of the sum in (12):
2
_
f
i
+f
i+1
2
_
2
tanh
_
x
i+1
x
i
2
_
2L
1
(x
i+1
x
i
) f
2
L

(xi,xi+1)
where L
1
= sup
z[0,1]
_
tanh(z/2)
z
_
and therefore
6 Holden and Raynaud
2
_
f
i
+f
i+1
2
_
2
tanh
_
x
i+1
x
i
2
_
6L
1
f
2
H
1
(xi,xi+1)
. (14)
As far as the other term in the sum is concerned, we have, by the Cauchy
Schwarz inequality, |f
i+1
f
i
|
_
xi+1
xi
|f
x
| dx

x
i+1
x
i
f
H
1
(xi,xi+1)
and therefore
2
_
f
i+1
f
i
2
_
2
coth
_
x
i+1
x
i
2
_

L
2
2
f
2
H
1
(xi,xi+1)
(15)
where L
2
= sup
z[0,1]
coth(z/2)z. Due to the Sobolev embedding H
1
(R)
L

(R), there exists a constant L

0
such that f
2
1
+ f
2
n
L
0
f
2
H
1
(R)
. Hence,
gathering (12), (14) and (15), we obtain (11). Let us prove now that for any
smooth function with compact support, I
P
() in H
1
(R) when
P
0
and L
P
. We denote v
P
= I
P
(). We have that v
P
satises in each
interval [x
i
, x
i+1
], v
P
v
P,xx
=
xx
and v
P
(x
i
) = v
P
(x
i+1
) = 0. By
integration by part, we obtain
v
P

2
H
1
(xi,xi+1)
=
_
xi+1
xi
(v
P
v
P,xx
)v
P
dx =
_
xi+1
xi
(
xx
)v
P
dx (16)

xx

L
2
(xi,xi+1)
v
P

L
2
(xi,xi+1)
,
after using CauchySchwarz. Hence,
v
P

H
1
(xi,xi+1)

xx

L
2
(xi,xi+1)
C
_

P
(17)
where C = sup
xR
(
xx
). Since v
P
(x
i
) = 0, we have v
P
(x)
2
= 2
_
x
xi
v
P
v
P,x
dx
2 v
P

2
H
1
(xi,xi+1)
and therefore, after using (17), we obtain v
P

(xi,xi+1)

2C

P
, which, together with (16), implies
v
P

2
H
1
(xi,xi+1)
2C
_

P
_
xi+1
xi
|
xx
| dx. (18)
When L
P
is large enough, and I
P
() vanish in (, x
1
) and (x
n
, ) so that
v
P

2
H
1
(R)
=

n1
i=1
v
P

2
H
1
(xi,xi+1)
. From (18), it follows that v
P

2
H
1
(R)

2C

P

xx

L
1
(R)
. Hence, we have proved that that I
P
() in H
1
(R)
when
P
0 and L
P
. We conclude the proof of the theorem by
a standard approximation argument. For any > 0, there exists a smooth
function with compact support such that f
H
1
(R)
. For any
P
1
small enough and L
P
large enough, we have I
P
()
H
1
(R)
. Hence, by
the uniform continuity of I
P
, we have
f I
P
(f)
H
1
(R)
f
H
1
(R)
+ I
P
()
H
1
(R)
+I
P
() I
P
(f)
H
1
(R)
(2 +C).
We choose a sequence of partition P
n
such that
Pn
0 and L
Pn
0.
We denote by x
i,n
the points of this partition, for example we can take x
i,n
=
i/n n for i = 1, . . . , n
2
. The numerical scheme described in the following
theorem converges.
Numerical scheme for the CamassaHolm equation 7
t=0 t=3
Fig. 1. Left: Approximation of the function u(x) = x(x
2
1)e
x
2
/4
by n = 10
equidistributed peakons. Right: Comparison of the computed solutions for n = 10
peakons (in bold) and n = 100 peakons at time t = 3.
Theorem 2. For any initial data u H
1
(R), let us dene u
n
M by u
n
=
I
Pn
( u). We denote by u
n
the solution of (8) with initial condition given by
( q
i,n
= x
i,n
, u
i,n
= u
n
(x
i,n
),

H
i,n
) where the

H
i,n
are computed from u
i,n
and
q
i,n
using (7). Then, u
n
converges to the global conservative solution u of the
CamassaHolm equation in C([0, T], L

(R)), for any T > 0.

Proof. By Theorem 1, we have that u
n
u in H
1
(R). From [9, Proposition
5.1] it follows that ( u
n
, ( u
2
n
+ u
2
x,n
) dx) converges to ( u
n
, ( u
2
+ u
2
x
) dx) in D. The
conservative solution of the CamassaHolm equation constitues a continuous
semigroup in D. Hence, ( u
n
, (u
2
n
+u
2
n,x
) dx) converges to (u, (u
2
+u
2
x
) dx) in
C([0, T], D) and therefore in C([0, T], L

(R)), by [9, Proposition 5.2].

4 Numerical example
We consider the initial data u given by u(x) = x(x
2
1)e
x
2
/4
. The com-
puted solution is shown in Figure 2 for dierent times t. Figure 1 shows that
the scheme is remarkably stable as, after a number of collisions, the solution
computed with only n = 10 peakons remains close to the solution computed
with much higher accuracy.
References
1. Alberto Bressan and Adrian Constantin. Global conservative solutions of the
CamassaHolm equation. Arch. Ration. Mech. Anal., to appear.
2. Alberto Bressan and Massimo Fonte. An optimal transportation metric for
solutions of the CamassaHolm equation. Methods Appl. Anal., 12:191220,
2005.
3. Roberto Camassa and Darryl D. Holm. An integrable shallow water equation
with peaked solitons. Phys. Rev. Lett., 71(11):16611664, 1993.
8 Holden and Raynaud
t=0 t=0.5 t=0.8
t=1 t=2 t=3
Fig. 2. The computed solution for n = 100 peakons shown at dierent times.
4. Roberto Camassa, Darryl D. Holm, and James Hyman. A new integrable shallow
water equation. Adv. Appl. Mech., 31:133, 1994.
5. G. M. Coclite, H. Holden, and K. H. Karlsen. Global weak solutions to a
generalized hyperelastic-rod wave equation. SIAM J. Math. Anal., 37(4):1044
1069 (electronic), 2005.
6. Giuseppe Maria Coclite, Helge Holden, and Kenneth Hvistendahl Karlsen.
Wellposedness for a parabolic-elliptic system. Discrete Contin. Dyn. Syst.,
13(3):659682, 2005.
7. Adrian Constantin. Existence of permanent and breaking waves for a shallow
water equation: a geometric approach. Ann. Inst. Fourier (Grenoble), 50(2):321
362, 2000.
8. Adrian Constantin and Joachim Escher. Wave breaking for nonlinear nonlocal
shallow water equations. Acta Math., 181(2):229243, 1998.
9. Helge Holden and Xavier Raynaud. Global conservative solutions of the
CamassaHolm equation a lagrangian point of view. Comm. Partial Dif-
ferential Equations, to appear.
10. Helge Holden and Xavier Raynaud. Global conservative solutions of the gener-
alized hyperelastic-rod wave equation. J. Dierential Equations, to appear.
11. Helge Holden and Xavier Raynaud. Convergence of a nite dierence scheme for
the camassaholm equation. SIAM Journal on Numerical Analysis, 44(4):1655
1680, 2006.
12. Helge Holden and Xavier Raynaud. A convergent numerical scheme for the
Camassa-Holm equation based on multipeakons. Discrete Contin. Dyn. Syst.,
14(3):505523, 2006.
13. Helge Holden and Xavier Raynaud. Global conservative multipeakon solutions
of the CamassaHolm equation. J. Hyperbolic Dier. Equ., to appear, 4(1):39
63, 2007.