Multi-Symplectic Runge-Kutta Collocation

Methods for Hamiltonian Wave Equations

Sebastian Reich
Department of Mathematics and Statistics
University of Surrey, Guildford, GU2 5XH
August 25, 1999

Abstract
A number of conservative PDEs, like various wave equations, allow for a multi-symplectic
formulation which can be viewed as a generalization of the symplectic structure of Hamiltonian
ODEs. We show that Gauss-Legendre collocation in space and time leads to multi-symplectic
integrators, i.e., to numerical methods that preserve a symplectic conservation law similar to
the conservation of symplecticity under a symplectic method for Hamiltonian ODEs. We also
discuss the issue of conservation of energy and momentum. Since time discretization by a
Gauss-Legendre method is computational rather expensive, we suggest several semi-explicit
multi-symplectic methods based on Gauss-Legendre collocation in space and explicit or linearly
implicit symplectic discretizations in time.

1 Introduction
The scalar wave equation
@tt u = @xx u ? V 0 (u); (x; t) 2 U  R 2 ; (1)
V : R ! R some smooth function, is an example of a multi-symplectic Hamiltonian PDE [3] of type
M@t z + K@xz = rz S (z) (2)
where z 2 R d , M; K 2 R dd are two skew-symmetric matrices (which can be singular), and S :
R d ! R is a smooth function. In the particular case of the wave equation (1), a multi-symplectic
formulation (2) can be derived as follows [3]. We introduce canonical momenta v = @t u, w = @xu
and dene the state variable 2u3
z = 4 v 5 2 R3:
w
Using this variable, we obtain
2 0 ?1 0
3 2 0 0 1
3
M=4 1 0 0 5 and K=4 0 0 0 5;
0 0 0 ?1 0 0
as well as the Hamiltonian ?

S (z) = 21 v2 ? w2 + V (u):

1
Other Hamiltonian PDEs that can be rewritten as a multi-symplectic system (2) include the non-
linear Schrodinger equation (see x6), the Boussinesq equation, the shallow water equation, and the
Korteweg-de Vries equation [3, 4, 5].
The multi-symplectic reformulation (2) is interesting for several reasons (see [3, 4, 5] for the-
oretical results), one of the most important perhaps being the existence of the multi-symplectic
conservation law
@t [dz ^ Mdz] + @x [dz ^ Kdz ] = 0 (3)
which, for the wave equation (1), is equivalent to
@t [du ^ dv] ? @x [du ^ dw] = 0: (4)
In other words, changes in the wedge product du ^ dv in time are exactly compensated for by changes
in the wedge product du ^ dw in space. We note that this is a strictly local conservation concept
that, e.g., does not depend on specic boundary conditions. This is in contrast to the classical
formulation of the wave equation (1) as a Hamiltonian initial value problem over some appropriate
function space. Furthermore, for conservative systems that do not depend on a spatial direction
(like Newton's equations of motion), we obtain the classical result
d [du ^ dv] = d X M
dt dt k=1 duk ^ dvk = 0
where u 2 R M would be the vector of coordinates and v 2 R M the vector of canonical momenta.
It is widely recognised that the symplectic structure of phase space should be taken into account
when it comes to numerical simulations of Hamiltonian systems. In particular, methods that preserve
the wedge product du ^ dv exist for canonical (nite-dimensional) Hamiltonian systems and are
called symplectic methods [18]. A discussion of symplectic methods for nite dimensional spatial
truncations of the wave equation (1) can be found, for example, in [13, 8]. The basic idea is to nd
a nite dimensional spatial truncation
d u = v; (5)
dt
d v = Au ? V 0 (u) (6)
dt
u = (u01; : : : ; uM 0)T , v =0(v1 ; : : : ; vM 0)T where A 2 R M M is a nite-dierence approximation to @xx
and V (u) = (V (u1 ); V (u2 ); : : : ; V (uM )) , M the number of grid points. This dierential equation
T
(5)-(6) is Hamiltonian in the variable (u; v) 2 R 2M if A is symmetric leading to the invariance of the
symplectic structure du ^ dv. Typically A is obtained using the second central dierence approx-
imation (leapfrog) but higher order symmetric nite dierence approximations are also occasionally
applied [8]. This system (5)-(6) can now be integrated in time using a symplectic discretization.
For example, if A is the second central dierence approximation and (5)-(6) is discretized by the
Stormer-Verlet method [18], then the popular leapfrog discretization of the nonlinear wave equation
(1) is obtained. Instead of applying a symplectic discretization in time one can also use an energy
conserving discretization. This option has, for example, been explored in [19, 10, 15].
Contrary to this \classical" approach, the investigation of multi-symplectic numerical methods
for Hamiltonian PDEs is a very recent subject. The rst paper is due to Marsden, Patrick &
Shkoller [11] who used a formulation of the multi-symplectic structure for wave equations based
on a Lagrangian formulation of the Cartan form in eld theory [12]. They then derive a numerical
scheme by discretizing the Lagrangian. This leads to an elegant generalization of Veselov discretiz-
ations [20] to PDEs. However, there are limitations in a Lagrangian setting, such as restrictions on
the order of numerical discretizations. A new denition of multi-symplectic integrators based on a

2
discrete form of the symplectic conservation law (3) was recently suggested by Bridges & Reich
in [6]. It has been shown that popular methods such as the centered Preissmann scheme [1] and
the leapfrog method are multi-symplectic and that such schemes have remarkable local energy and
momentum conserving properties [6].
In this paper, we like to explore the question of whether or not symplectic methods
zn+1 =
t(zn); tn+1 = tn +
t;
for canonical Hamiltonian ODEs
d z = J?1 r H (z); z 2 R 2M ; J = ?JT ; (7)
dt z

with their exact conservation property

dzn+1 ^ Jdzn+1 = dzn ^ Jdzn
can be generalised to multi-symplectic PDEs (2). The main objective is to formulate higher order
multi-symplectic discretizations. For simplicity of exposition, we will restrict to the nonlinear wave
equation (1) and one of our main results is that we show that a proper implementation of Gauss-
Legendre collocation methods [18] leads to a multi-symplectic numerical scheme for (1). See x6 for a
brief discussion of the non-linear Schrodinger equation. We point out that the spatial discretization
by Gauss-Legendre collocation methods leads to a particular type of the pseudospectral (cell) meth-
ods [9]. It is known that these methods are well suited for the integration of very smooth solutions
over long periods of time under a stringent error requirement [9].
One of the main justications for applying symplectic methods to (nite-dimensional) Hamilto-
nian systems (7) stems from their excellent conservation of the Hamiltonian (energy) H over very
long integration intervals tn 2 [0; T ] [2]. For a multi-symplectic system (2), conservation of energy
is replaced by the following local conservation law [4]:
@t E (z ) + @x F (z ) = 0: (8)
The energy density is
E (z ) = S (z ) ? 21 zT K@xz
and the energy ux is
F (z ) = 21 z T K@t z :
This can be seen from
0 = @t zT M@t z = @t S (z ) ? @t zT K@xz = @t S (z) ? 21 @t [z T K@x z] + 12 @x [z T K@t z ]:
For the wave equation (1), we can eliminate the partial derivatives in E and F and can formulate
an energy conservation law (8) with
?

E = 12 w2 + v2 + V (u) and F = ?vw
where we have made use of the identity @x v = @t w = @tx u. There is a corresponding momentum
conservation law which is obtained by premultiplying (2) by @x zT [4]. For the wave equation, we
can again simplify the corresponding expressions and obtain a momentum conservation law
@t I (z) + @x G(z ) = 0 (9)

3
with ?

I (z) = F (z ) = ?vw and G(z ) = 21 w2 + v2 ? V (u):
It is known that Gauss-Legendre methods exactly preserve quadratic rst integrals of ordinary
dierential equations [7, 18]. One could therefore expect that a Gauss-Legendre discretization of the
wave equation (1) leads to a discrete energy/momentum conservation law if the function V is zero
or at most quadratic in u. As we will see in x3, this analogy does indeed hold.

2 Gauss-Legendre Discretizations are Multi-Symplectic

Let us state the multi-symplectic formulation of the wave equation (1) in the form
@t u = v; (10)
@x u = w; (11)
0
@t v ? @x w = ?V (u): (12)
The corresponding equations for the dierential one forms dz = (du; dv; dw) are given by
T
@t du = dv;
@x du = dw;
@t dv ? @x dw = ?V 00 (u)du:
This yields the multi-symplectic conservation law (4) via
du ^ @tdv ? du ^ @xdw = ?V 00 (u)du ^ du = 0
together with the observation that
@t [du ^ dv] = @t du ^ dv + du ^ @t dv = dv ^ dv + du ^ @t dv = du ^ @t dv
and, similarly,
@x [du ^ dw] = du ^ @x dw:
Let us now discretize (10)-(12) by a Gauss-Legendre collocation method written in the form of
an implicit Runge-Kutta method [18]. We start with the spatial discretization and rewrite (10)-(12)
as
@x u = w; (13)
0
@x w = @t v + V (u); (14)
0 = @t u ? v (15)
Next we apply an implicit Runge-Kutta (RK) scheme with s stage variables (Ui ; Wi ) and coecients
faij g, fbi g to (13)-(14) and obtain the semi-discretization
X
s
Ui = uk +
x aij Wj ; (16)
j =1
X s
Wi = wk +
x aij (@t Vj + V 0 (Uj )) ; (17)
j =1
X s
uk+1 = uk +
x bi Wi ; (18)
i=1
X s
wk+1 = wk +
x bi (@t Vi + V 0 (Ui )) (19)
i=1

4
which is dened for all t. Note that uk (t) is an approximation to u(xk ; t). For simplicity, we set
k = 0 and assume that xk = 0.
Let us check if this semi-discretization of the wave equation (2) implies a corresponding semi-
discretized version of the conservation law (4). We rst derive the equations for the dierentials
X
s
dUi = du0 +
x aij dWj ; (20)
j =1
X s
dWi = dw0 +
x aij (@t dVj + V 00 dUj ) ; (21)
j =1
X s
du1 = du0 +
x bi dWi ; (22)
i=1
X s
dw1 = dw0 +
x bi (@t dVi + V 00 dUi ) : (23)
i=1
Using (22) and (23), we proceed with
X
s ! X
s !
du1 ^ dw1 = du0 +
x bi dWi ^ dw0 +
x bi (@t dVi + V 00 dUi )
i=1 i=1
X
s X s
= du0 ^ dw0 + du0 ^
x bi (@t dVi + V 00 dUi ) +
x bi dWi ^ dw0 +
i=1 i=1
X
s X s
+
x2 bi bj dWi ^ (@t dVj + V 00 dUj ) :
i=1 j =1
This is further transformed to
X s
du1 ^ dw1 = du0 ^ dw0 +
x bidUi ^ @t dVi +
i=1
X
s X s
+
x2 (bi bj ? bj aji ? bi aij ) dWi ^ (@t dVj + V 00 dUj )
i=1 j =1
using (20)-(21). The last term is zero for Gauss-Legendre methods since [18]
bj aji + bi aij ? bi bj = 0
for all i; j . Thus, upon applying a Gauss-Legendre RK method, we obtain the semi-discretized
conservation law
Xs
[du1 ^ dw1 ? du0 ^ dw0 ] ? bi [dUi ^ @t dVi ]
x = 0 (24)
i=1
which holds for all t.
The next step is the discretization in time over a time interval [0;
t]. Since we are using an
implicit RK methods, we can solve (16)-(17) for @t Vj (t), j = 1; : : : ; s. We also use equation (15) and
implement a Runge-Kutta time-discretization with r stages and coecients fa~mng, f~bm g:
X
r
Ui;m = u0i +
t a~mn Vi;n ; (25)
n=1

5
 X
r
Vi;m = vi0 +
t a~mn @t Vi;n ; (26)
n=1
Xr
u1i = u0i +
t ~bm Vi;m ; (27)
m=1
Xr
vi1 = vi0 +
t ~bm @t Vi;m : (28)
m=1
Here we introduced the notation
Ui;m  u(ci
x; dm
t); u1i  u(ci
x;
t); um1  u(
x; dm
t)
etc. with
X
s X
r
ci = aij and dm = a~mn :
j =1 n=1
The symbol @t Vi;m denotes the numerical approximation of @t Vi (dm
t) which we obtain from (16)-
(17) by xing t = dm
t.
Now we assume again that a Gauss-Legendre method is used and, upon applying similar argu-
ments as before, we obtain the identity
X r ! X r !
du1i ^ dvi1 = du0i +
t ~bmdVi;m ^ dvi0 +
t ~bm@tdVi;m
m=1 m=1
Xr
= du0i ^ dvi0 +
t ~bm dUi;m ^ @t dVi;m
m=1
or
du1 ^ dv1 ? du0 ^ dv0  ? X
r
~bm [dUi;m ^ @t dVi;m ]
t = 0:
i i i i
m=1
Next we rewrite (24) for t = dm
t as
X
s
[dum
1 ^ dw1m ? dum
0 ^ dw0m ] ? bi [dUi;m ^ @t dVi;m ]
x = 0
i=1
and obtain the discretized conservation law
X
s   X
r
bi du1i ^ dvi1 ? du0i ^ dvi0
x ? ~bm [dum1 ^ dw1m ? dum0 ^ dw0m ]
t = 0 (29)
i=1 m=1
This formula can be understood as the approximation of the integral of (4) over the domain of
integration [0;
x]  [0;
t], i.e., (29) approximates
Z
x
[du(x;
t) ^ dv(x;
t) ? du(x; 0) ^ dv(x; 0)] dx ?
x=0
Z
t
? [du(
x; t) ^ dw(
x; t) ? du(0; t) ^ dw(0; t)] dt = 0 (30)
t=0
using Gaussian quadrature for the evaluation of the two integrals.

6
Proposition 1. Let the multi-symplectic formulation (10)-(12) of the nonlinear wave equation
(1) be discretized in space and in time by a pair of Gauss-Legendre collocation methods with s, r
respectively, stages. Then the resulting discretization is a multi-symplectic integrator with discrete
conservation law (29).
It might appear as if the order in which the discretization in space and time is carried out would
lead to dierent schemes. But this is not the case. Take, for example, the spatial discretization
(16)-(19), which we write in its time-sampled version
X
s
Ui;m = um0 +
x aij Wj;m ; (31)
j =1
X s
Wi;m = w0m +
x aij (@t Vj;m + V 0 (Uj;m )) ; (32)
j =1
Xs
um1 = um0 +
x bi Wi;m ; (33)
i=1
Xs
w1m = w0 +
x bi (@t Vi;m + V 0 (Ui;m )) ; (34)
i=1
and the time discretization (25)-(28). We could apply the substitution
@x Wi;m = @t Vi;m + V 0 (Ui;m )
in (32) and (34) as well as
@t Vi;m = @x Wi;m ? V 0 (Ui;m )
in (26) and (28) to obtain the same scheme which, however, appears now to be obtained via a semi-
discretization in time followed by a discretization in space. In fact, our schemes can be written in
the general form as presented in Fig. 1.
We like to point out that s = r = 1, i.e., an implicit midpoint discretization in space and time,
leads to the centered Preissmann scheme1 [1, 6]
1 M z1 ? z 0  + 1 K z1=2 ? z 1=2  = r S (z 1=2 ) (35)

t 1=2 1=2
x 1 0 z 1=2
when applied to a general multi-symplectic PDE (2). Here the standard abbreviations
z01=2 = 21 ?z00 + z01
; z10=2 = 12 ?z00 + z10
; z11==22 = 14 ?z11 + z01 + z10 + z00
etc.
are used. The corresponding discrete multi-symplectic conservation law is [6]
h 1 i h i
dz1=2 ^ Mdz11=2 ? dz01=2 ^ Mdz01=2
x + dz11=2 ^ Kdz11=2 ? dz10=2 ^ Kdz10=2
t = 0:
See [6] for a detailed discussion of the Preissmann scheme and its application to multi-symplectic
PDEs (2).
When looking at problems with periodic boundary conditions, formula (29) has another interest-
ing consequence. Let us take the sum of the equation (29) over all spatial grid points k = 1; : : : ; M
X
M X
s  X !
bi du1i;k ^ dvi;k
1 ? du0i;k ^ dvi;k 
0
x ?
r 
~bm dum 
k+1 ^ dwk+1 ? duk ^ dwk
t = 0:
m m m
k=1 i=1 m=1
1 Preissmann originally suggested the method in 1961 for the solution of one dimensional open-channel hydraulics
problems. The scheme is a popular choice for the computation of unsteady ow problems. It is less well-known that
the scheme can also be applied to wave equations (see [1] for a historical account).

7
 X
s
Ui;m = um0 +
x aij @x Uj;m ;
j =1
X s
Wi;m = w0m +
x aij @x Wj;m ;
j =1
Xs
um1 = um0 +
x bi @x Ui;m ;
i=1
Xs
w1m = w0 +
x bi @x Wi;m ;
i=1

X
r
Ui;m = u0i +
t a~mn @t Ui;n ;
n=1
Xr
Vi;m = vi0 +
t a~mn @t Vi;n ;
n=1
Xr
u1i = u0i +
t ~bm @t Ui;m ;
m=1
Xr
vi1 = vi0 +
t ~bm @t Vi;m ;
m=1

@t Ui;m = Vi;m ;
@x Ui;m = Wi;m ;
@t Vi;m ? @x Wi;m = ?V 0 (Ui;m ):

Figure 1: Multi-Symplectic Gauss-Legendre Collocation Methods for Nonlinear Wave Equation.

Here we have used the abbreviations

u1i;k  u(ci
x + k
x;
t); umk  u(k
x; dm
t)
etc. Periodicity in space implies
X
M X
r
~bm dum m ? dum ^ dwm 
t = X ~bm dum ^ dwm ? dum ^ dwm 
t = 0
r
k+1 ^ dw k+1 k k M +1 M +1 1 1
k=1 m=1 m=1
which in turn yields
X
M X
s X
M X
s
bi du1i;k ^ dvi;k
1 = bi du0i;k ^ dvi;k
0 :
k=1 i=1 k=1 i=1
This is precisely conservation of symplecticity in time with respect to the state variables
u = fui;k gk=1;:::M ;i=1;:::;s and v = fvi;k gk=1;:::M ;i=1;:::;s
8
and the wedge product du ^ Bdv where B is a diagonal matrix with entries corresponding to the
weights fbi g. In other words, multi-symplectic Gauss-Legendre methods imply the conservation of
symplecticity in time, i.e.
dun+1 ^ Bdvn+1 = dun ^ Bdvn;
when applied to problems with periodic boundary conditions. Thus these methods are also sym-
plectic discretizations in the \classical" sense (although with a non-canonical - but non-degenerate
- symplectic structure).
Proposition 2. Let the multi-symplectic formulation (10)-(12) of the nonlinear wave equation
(1) be discretized in space and in time by a pair of Gauss-Legendre collocation methods with s, r
respectively, stages. Assume that we solve an initial value problem in time with periodic boundary
conditions in space. Then the resulting discretization yields a nite dimensional Hamiltonian
truncation of the wave equation in space with the underlying symplectic structure du ^ Bdv and a
symplectic discretization of this nite-dimensional system in time.

3 Conservation of Energy and Momentum

As already mentioned in the Introduction, the wave equation (1) satises the energy conservation
law
@t E + @ x F = 0
where ?

E = 21 w2 + v2 + V (u)
is the energy density and
F = ?vw
is the energy ux. The wave equation also satises the momentum conservation law (9). We will
come back to the issue of momentum conservation towards the end of this section.
We now check for the conservation of the following discrete energy conservation law
X
s   X
r
bi Ei1 ? Ei0
x + ~bm [F1m ? F0m ]
t = 0 (36)
i=1 m=1
with the obvious abbreviations
?
?

Ei1 = 21 wi1 2 + vi1 2 + V (u1i ); F1m = ?v1m w1m ;
etc. This is an approximation to the contour integral in the (x; t) plane
Z k
x Z n
t
[E (z (x; n
t)) ? E (z (x; 0))] dx + [F (z (k
x; t)) ? F (z(0; t))] dt = 0 (37)
x=0 t=0
with k = n = 1 in the same sense as (29) is for (30). It immediately also implies a corresponding
result for any k; n > 1.
Since Gauss-Legendre collocation methods preserve quadratic integrals of ODEs exactly [7, 18],
we could expect that multi-symplectic Gauss-Legendre methods will give rise to a discrete conser-
vation law (36) when V (u) = cu2=2, c 2 R some constant. To be able to evaluate the corresponding

9
discrete conservation law (36), we have to rst introduce two auxiliary systems, namely
X
s
Vi;m = v0m +
x aij @x Vj;m ; (38)
j =1
X s
v1m = v0m +
x bi @x Vi;m (39)
i=1
and
X
r
Wi;m = wi0 +
t a~mn @t Wi;n ; (40)
n=1
Xr
wi1 = wi0 +
t ~bm @t Wi;m (41)
m=1
to obtain the missing approximations of v and w along the boundary of [0;
x]  [0;
t]. The
undetermined values of fv0m g and fwi0 g are implicitly given by
X
r
um0 = u00 +
t a~mn v0n (42)
n=1
and
X
s
u0i = u00 +
x aij wj0 : (43)
j =1
Using the equations (38)-(41), (25), and (31), we derive the two identities
X
r
Ui;m = u0i +
t a~mn Vi;n
n=1 0 1
Xr X
s
= u0i +
t a~mn @v0n +
x aij @x Vj;n A
n=1 j =1
Xr X s Xr
= u0i +
t a~mn v0n +
t
xaij a~mn @x Vj;n
n=1 j =1 n=1
0 m 0 X s X r
= ui + u0 ? u0 +
t
x aij a~mn @x Vj;n
j =1 n=1
and, similarly,
X
s X
r
Ui;m = um0 + u0i ? u00 +
t
x aij a~mn @t Wj;n :
j =1 n=1
This implies
X
s X
r X
s X
r
aij a~mn @x Vj;n = aij a~mn @t Wj;n
j =1 n=1 j =1 n=1
and we can conclude that
@x Vi;m = @t Wi;m for all i; m: (44)
This re ects the fact that the exact solutions also satisfy @x v = @t w = @xt u. (In [6], this \hidden"
conservation law is made explicit by going to a dierent multi-symplectic formulation of the wave
equation (1).)

10
 Note that we also obtain the additional equations
X
r X
s
u10 = u00 +
t ~ bm v0m and u01 = u00 +
x bi wi0
m=1 i=1
as well as
X
r X
s
u11 = u01 +
t ~ bmv1m and u11 = u10 +
x bi wi1 :
m=1 i=1
0 0 1 1
which dene u0  u(0; 0), u1  u(
x; 0), u0  u(0;
t), and u1  u(
x;
t) and thus close the
auxiliary system (38)-(43). These equations are over-determined but self-consistent because of
X
r Xs X r Xs
0 =
t bmv1 +
x bi wi ?
t bmv0 ?
x bi wi1
~
m 0 ~ m
m=1 i=1 m=1 i=1
Xr Xr X s Xs
~bm bi @x Vi;m +
x bi wi0 ?
=
t ~bmv0m +
t
x
m=1 m=1 i=1 i=1
Xr X s Xs X r
?
t ~bm v0m ?
x bi wi0 ?
x
t bi~bm@t Wi;m
m=1 i=1 i=1 m=1
and (44). In fact, we have just shown that the \hidden" conservation law vt ? wx = 0 gives rise to
the discrete conservation law
X
r X
s  
~ bm [v1m ? v0m ]
t ? bi wi1 ? wi0
x = 0:
m=1 i=1
Let us now come back to the discrete energy conservation law (36). The equations (38)-(41)
together with (31)-(34) and (25)-(28) yield
X s ! Xs !
w1m v1m = m m
w0 +
x bi (@t Vi;m + cUi;m ) v0 +
x bi @x Vi;m
i=1 i=1
X s
= w0m v0m +
x bi (Vi;m @t Vi;m + cVi;m Ui;m + Wi;m @x Vi;m ) (45)
i=1
and, similarly,
X r
Ei1 = Ei0 +
t ~bm (Vi;m @t Vi;m + cUi;m Vi;m + Wi;m @t Wi;m ) (46)
m=1
where we used the fact that
bj aji + bi aij ? bi bj = 0 and ~bj a~ji + ~bi a~ij ? ~bi~bj = 0
in the same manner as done in x2 to prove multi-symplecticity. Combining both formulas, we obtain
X
s   X
r X
s X
r
bi Ei1 ? Ei0
x + ~bm [F1m ? F0m ]
t = bi~bm Wi;m (@t Wi;m ? @x Vi;m ) :
i=1 m=1 i=1 m=1
The right hand side of this equation is zero due to (44).

11
Proposition 3. Let the multi-symplectic formulation (10)-(12) of the nonlinear wave equation
(1) be discretized in space and in time by a pair of Gauss-Legendre collocation methods with s, r
respectively, stages. Let us also assume that the potential V in (1) is of the form
V (u) = 2c u2;
c 2 R some constant. Then the resulting discretization conserves the discrete energy conservation
law (36).
Li & Vu-Quoc have suggested a method that exactly conserves a discrete local energy conservation
law for any type of smooth potential V (u) [21, 15]. This local energy conservation law is, however,
dierent from (36). Their approach is based on a second central dierence approximation in space
and an energy conserving variant of the implicit midpoint method in time. We wish to point out
that, following their approach, one can also obtain a variant of the centered Preissmann scheme
(35) that exactly conserves the discrete energy conservation law (36) with s = r = 1 for any smooth
potential V (u). The idea is to apply an implicit midpoint discretization in space and a proper energy
conserving variant of the midpoint discretization in time. See the Appendix for more details. Note
that this energy conserving scheme cannot be expected to be multi-symplectic, in general, and that,
for linear problems, the scheme reduces to the box scheme.
We now investigate the preservation of the momentum conservation law (9). We rst dene a
discrete conservation law
X
s   X
r
bi Ii1 ? Ii0
x + ~bm [Gm1 ? Gm0 ]
t = 0 (47)
i=1 m=1
with the obvious abbreviations
 
Ii1 = ?wi1 vi1 ; Gm1 = 21 (v1m )2 + (w1m )2 ? V (um1 );
etc. This is an approximation to the contour integral in the (x; t) plane,
Z
x Z
t
[I (z (x;
t)) ? I (z (x; 0))] dx + [G(z (
x; t)) ? G(z (0; t))] dt = 0:
x=0 t=0
Assuming V (u) = cu2=2, we rst derive
X r ! X r !
wi1 vi1 = wi0 +
t ~bm@t Wi;m vi0 +
t ~bm @t Vi;m
m=1 m=1
Xr
= wi0 vi0 +
t ~bm (Vi;m @x Vi;m + Wi;m @t Vi;m )
m=1
Xr
= wi0 vi0 +
t ~bm (Vi;m @x Vi;m + Wi;m @xWi;m ? cWi;m Ui;m )
m=1
where, in the second line, we made use of (44). We also obtain
m m Xs
G1 = G0 +
x bi (Wi;m @x Wi;m + Vi;m @x Vi;m ? cUi;m Wi;m )
i=1
Combining both results, we indeed obtain the discrete conservation law (47).

12
Proposition 4. Let the multi-symplectic formulation (10)-(12) of the nonlinear wave equation
(1) be discretized in space and in time by a pair of Gauss-Legendre collocation methods with s, r
respectively, stages. Let us also assume that the potential V in (1) is of the form
V (u) = 2c u2;
c 2 R some constant. Then the resulting discretization conserves the discrete momentum conserva-
tion law (47).
Proposition 3 & 4 show that multi-symplectic integrators for linear wave equations have excellent
energy-momentum conservation. This suggests that multi-symplectic integrators for nonlinear wave
equations will preserve the discrete energy-momentum conservation laws in good approximation.
The following \global" result for a wave equation (1) with periodic boundary conditions u(0; t) =
u(L; t) can be stated. It follows from (37) that
ZL ZL
E (t) := E (z (x; t))dx = E (0) := E (z (x; 0))dx
x=0 x=0
if z (0; t) = z (L; t) (periodic boundary conditions). This implies conservation of total energy E (t)
for the wave equation when considered as an innite-dimensional Hamiltonian PDE. Any nite-
dimensional Hamiltonian approximation of the wave equation and subsequent symplectic integration
in time will conserve this energy very well provided
t small enough [2]. This also applies to the
multi-symplectic Gauss-Legendre methods of x2 which follows from Proposition 2 and backward
error analysis results in [2, 16]. Furthermore, there is also conservation of global momentum
ZL ZL
I (t) := I (z (x; t))dx = I (0) := I (z (x; 0))dx
x=0 x=0
if z(0; t) = z(L; t) (periodic boundary conditions). Since I is a quadratic expression in v and w,
Gauss-Legendre collocation methods will preserve the discrete global momentum conservation law
exactly.
However, this global view provides a much weaker concept of energy and momentum conserva-
tion. It is a necessary but not sucient condition for the preservation of the corresponding local
conservation laws.

4 Semi-Explicit and Linearly Implicit Multi-Symplectic Dis-

cretizations
There are certainly other options available for the time discretization then using a high-order implicit
Gauss-Legendre collocation method (25)-(28) when integrating an initial-boundary value (Cauchy)
problem. In fact, one can either apply the implicit midpoint method in time and use appropriate
concatenations of the corresponding scheme to obtain a higher order in time method or one can
apply a higher order explicit symplectic partitioned Runge-Kutta method in time [18]. Let us, for
example, take an explicit symplectic concatenation method of the form
Ui;1 = u0i ;
Vi;1 = vi0 ;
for m = 1; : : : ; r Ui;m+1 = Ui;m + m
tVi;m+1 ;
Vi;m+1 = Vi;m + m
t@t Vi;m ;
1
ui = Ui;r+1;
u1i = Vi;r+1

13
with weights (m ; m ). This yields
du1 ^ dv1 ? du0 ^ dv0 ? X
r
m [dUi;m ^ @t dVi;m ]
t = 0
i i i i
m=1
which in turn implies the discrete conservation law
X
s   X
r
bi du1i ^ dvi1 ? du0i ^ dvi0
x ? m [dum1 ^ dw1m ? dum0 ^ dw0m ]
t = 0
i=1 m=1
which is again a discretization of (30).
Another option is to apply the multi-symplectic Gauss-Legendre collocation methods from Fig. 1
to a linearized wave equation and to use the following three step discretization for the nonlinear
problem:
1. Update vi0 via
v~i0 = vi0 ?
2t V 0 (u0i ) :
2. Apply a multi-symplectic Gauss-Legendre method to the linearized wave equation using v~i0 .
1 ; w1m ; u1i ; v~i1 )T .
Denote the result by (um
3. Update v~i1 via
vi1 = v~i1 ?
2t V 0 (u1i ) :
Here V (u) = cu2 =2 + V (u) was used. This linearly implicit scheme conserves the multi-symplectic
conservation law (29). The overall scheme will only be second order in time, in general. But higher
order in time schemes can be obtained by a proper composition of the (symmetric) second order
scheme [14].

5 Numerical Experiments
To gain insight into the performance of the suggested multi-symplectic methods, in particular into
their local conservation of energy and momentum, we perform the following numerical experiments.
We discretize the sine-Gordon equation
@tt u = @xx u ? sin(u)
using the multi-symplectic Gauss-Legendre method with r = 1 (midpoint in time) and s = 2 (fourth
order Gauss-Legendre method in space). We implement periodic boundary conditions u(?L=2; t) =
u(L=2; t) for two dierent sets of initial conditions. For the rst experiment, which we call Experiment
A, we use L = 100,
x?L=6
! !
u(x; 0) = 4 tan?1 e?x?L=6 ;
pe 1 ? 2 + 4 tan?1 p 1 ? 2
and " x?L=6?t
! ?x?L=6?t
!#
@ 4 tan?1 ep
v(x; 0) = @t + 4 tan?1 e p :
1 ? 2 1 ? 2 t=0
On an innite domain, these initial conditions correspond to a soliton and anti-soliton solution
moving with speed  . We set  = 0:5. In Fig. 2, snapshots of the time evolution of the wave form

14
 time t=0 time t=17

5 5

0 0

5 5

10 10

50 0 50 50 0 50
time t=67 time t=117
5
5

0 0

u(x)

u(x)
5
5
10
10
50 0 50 50 0 50
time t=167 time t=200

5 5

0 0

5 5

10 10

50 0 50 50 0 50
x x

Figure 2: Time evolution of the soliton/anti-soliton solution over the time interval t 2 [0; 200].

u(x; t) are shown over a time interval t 2 [0; 200]. This approximation is obtained using a time-step
of
t = 0:1 and M = 100 spatial mesh points.
For the second experiment, called Experiment B, we consider the so-called breather solution
p !
u(x; t) = 4 tan?1 1 ? !2 cosp!t
! cosh (x 1 ? !2 )
with ! = 0:9. On an innite domain, this is a bump shaped solution which oscillates up and down
with period 2=!. To exclude boundary eects, we used period boundary conditions with L = 100.
We monitor the drift in the numerical approximations to the total energy E (t) and the total
momentum I (t) as well as the error in the numerical discretizations of the local energy conservation
law (36) and the local momentum conservation law (47).
The numerical errors in the conserved quantities for the Experiment A can be found in Fig. 3.
These results are obtained using M = 100 spatial grid points and a time-step of
t = 0:1. Total
momentum I (t) is conserved exactly since it is a quadratic invariant. The two pronounced spikes in
the error of the total energy correspond to rapid \ ip" transitions of the solution u(x; t) whenever
the two kinks meet at x = L or x = 0.
We also plot the error in the local conservation laws (36) and (47) as a function of the spatial
grid location and the time step. As can be seen from Fig. 4, the errors are mainly concentrated
around the two moving fronts.
The corresponding numerical results for the Experiment B can be found in Fig. 5. These results
are obtained using M = 60 spatial grid points and 60 time steps per period of the motion which
is T = 2=!  6:9813. Again, total momentum I (t) is conserved exactly (up to the error in the
Newton iteration) since it is a quadratic invariant. The error behaviour of the total energy is typical
for a symplectic time discretization of an Hamiltonian ODE. The errors in the local conservation laws
are focused about x = 0 where the center of the breather solution is located. Instead of summing
the errors in the local conservation laws over the full spatial extension and plotting the result as a
function of time, one can also sum the errors over a xed spatial interval [0; a] and time intervals
[0;  ],   0. See Fig. 6 for the case  2 [0; 140] and a = L=4. Note that the errors in the local

15
 3 error in total energy
x 10
4

4
0 20 40 60 80 100 120 140 160 180 200
time

9 error in total momentum

x 10
1

0.5

0.5

1
0 20 40 60 80 100 120 140 160 180 200
time

Figure 3: Numerical errors in the global energy and momentum over the time interval [0; 200] (with
two collisions of kinks).
motion of two kinks motion of two kinks
5 3
x 10 x 10

5 error in momentum CL 2
error in energy CL

0 0

5 2
30 40
20 50 50
20
10 0 0
time 0 50 space time 0 50 space

collision of two kinks collision of two kinks

3 3
x 10 x 10
error in momentum CL

2 2
error in energy CL

1
0
0

1 2
170 170
50 50
160 160
0 0
time 150 50 space time 150 50 space

Figure 4: Numerical errors in the local energy/momentum conservation laws over the two intervals
[0; 30] (no collision of kinks) and [150; 170] (collision of kinks at x = 0).

energy conservation law do not grow with  but that the corresponding errors in the momentum
conservation laws increase with increasing values of  .
We like to point out that our numerical experiments show that the error in the energy conservation
law (36) divided by the area element
x
t depends only on the chosen time-step
t and is second
order in
t while the corresponding error in the local momentum conservation law (47) depends
only on the spatial mesh-size
x and is approximately of fourth order in
x. See Table 5.

16
 4 4
x 10 x 10

error in momentum CL
1

error in energy CL
2

0 0

2
1

100 50 100 50
50 0 50 0
time 0 50 space time 0 50 space

4 11
x 10 x 10

5 5

total error in momentum

error in total energy
0 0

5 5

0 50 100 0 50 100
time time

Figure 5: Numerical errors in the local/global energy/momentum conservation laws for the breather
solution over the time interval [0; 140].
3 (a) residual in the energy conservation law
x 10
3.5

2.5

1.5

0.5

0
0 20 40 60 80 100 120 140
time

3 (a) residual in the momentum conservation law

x 10
0

6
0 20 40 60 80 100 120 140
time

Figure 6: Numerical errors in the local energy (a) and momentum (b) conservation laws summed
over the spatial interval [0; L=4] and time intervals [0;  ] as a function of  2 [0; 140] (breather
solution).

For comparison, we implemented a method-of-line approach to the wave equation (1) using the
Stormer-Verlet method [18] in time and the symmetric fourth order nite dierences discretization
of the Laplacian operator [8]. The resulting discretization can be written as
qn+1 = qn +
tpn+1=2 ;
j j j

17
 Energy/Momentum
t = 0:058
t = 0:116
t = 0:1745

x = 1:0 3.2e-4/4.16e-5 1.3e-3 4.09e-5 2.8e-3/4.06e-5

x = 1:25 3.03e-4/1.06e-4 1.2e-3/1.06e-4 2.7e-3/1.04e-4

x = 1:67 3.06e-4/2.72e-4 1.1e-3/2.97e-4 2.4e-3/2.87e-4
Figure 7: Maximum error in the local energy/momentum conservation laws over all spatial grid
points and two time periods as a function of the spatial mesh size
x and the time step
t. The
residuals in (36) and (47) are divided by
t
x to make the corresponding integrals independent of
the size of the cell over which they are taken.
(a) finite differences (o), GaussLegendre (x)
1
10

2
global error 10

3
10

4
10
0 100 200 300 400 500 600 700
number of periods

(b) finite differences (o), GaussLegendre (x)

0
10

1
10
global error

2
10

3
10
0 100 200 300 400 500 600 700
number of periods

Figure 8: Global solution error for the breather solution after k = 1; : : : ; 700 periods for
t = T=120
and dierent values of
x = L=M using a symplectic nite dierences method and a multi-symplectic
Gauss-Legendre collocation method of the same order: (a) Gauss-Legendre method (x) with M = 100
and nite dierences method (o) with M = 150 mesh points; (b) Gauss-Legendre method (x) with
M = 80 and nite dierences method (o) with M = 100 mesh points.
 ?qn + 16qn ? 30qn + 16qn ? qn 
pnj +1=2 = pnj ?1=2 +

xt2 j +2 j +1
12
j j ?1 j ?2 ?
tV 0 (q n ) :
j

As shown in [13, 8], the resulting method is symplectic when considered as a time discretization
of a nite dimensional Hamiltonian approximation of (1). However, because of the non-symplectic
nite dierence approximation of the Laplacian, the overall method is not multi-symplectic. We
repeated the Experiment B with this nite dierences scheme and evaluated the l1 global error of
the numerical computed breather solution after k periods, k = 1; : : : ; 700. The results for a step-size
of
t = T=120, T = 2=!, and dierent values of the spatial mesh-size
x = L=M , L = 100, can
be found in Fig. 8. We note that the symplectic nite dierences method has initially a smaller
global error compared to the multi-symplectic Gauss-Legendre collocation method. However, due to
an apparently more favourable error propagation, the multi-symplectic collocation method yields,
in both cases, a smaller global error after k = 700 periods. Note that both methods use a second
order discretization in time and a fourth order discretization in space.

18
6 Another Example: Nonlinear Schrodinger Equation
To demonstrate the generality of our approach, we brie y consider multi-symplectic integration of
the one dimensional nonlinear Schrodinger equation [5]
i@t = @xx + V 0 (j j2 ) ; (x; t) 2 U  R 2 ;
V : R ! R some smooth function. Using = a + ib, we rewrite this as a pair of real-valued equations
@t a = +@xxb + V 0 (a2 + b2 )b;
@t b = ?@xxa ? V 0 (a2 + b2 )a:
Next we introduce a pair of conjugate momenta v = ax, w = bx , and obtain the multi-symplectic
PDE
?@t b ? @x v = V 0 (a2 + b2 )a;
@t a ? @x w = V 0 (a2 + b2 )b;
@x a = v;
@x b = w
with phase space variable z = (a; b; v; w)T and Hamiltonian
?

S (z ) = 21 v2 + w2 + V (a2 + b2 ) :
The corresponding multi-symplectic conservation law (3) is
@t [da ^ db] + @x [da ^ dv + db ^ dw] = 0: (48)
We like to point out that one of the conservation laws of the Schrodinger equation is
2 2
@t a +2 b + @x (vb ? wa) = 0 : (49)
Again we rst discretize the reformulated equation
@x v = ?@t b ? V 0 (a2 + b2 )a;
@x w = +@t a ? V 0 (a2 + b2 )b;
@x a = v;
@x b = w
in space by a Gauss-Legendre collocation method. Using similar notations as in x2, it is relatively
straightforward to derive the corresponding semi-discretized conservation law
X
s
[da1 ^ dv1 ? da0 ^ dv0 ] + [db1 ^ dw1 ? db0 ^ dw0 ] + bi @t [dAi ^ dBi ]
x = 0:
i=1
Discretization in time is now achieved by using
0 Xr
Ai;m = ai +
t a~mn @t Ai;n ;
n=1
Xr
Bi;m = b0i +
t a~mn @t Bi;n ;
n=1
X
r
a1i = a0i +
t ~ bm@t Ai;m ;
m=1
Xr
b1i = b0i +
t ~bm @t Bi;m
m=1

19
with corresponding conservation property
da1 ^ db1 ? da0 ^ db0 ? X
r
~bm [@t dAi;m ^ dBi;m + dAi;m ^ @t dBi;m ]
t = 0:
i i i i
m=1
Combining these results, we derive the discretized multi-symplectic conservation law
Xs  
bi da1i ^ db1i ? da0i ^ db0i
x +
i=1
Xm
+ ~bm [dam 1 ^ dv1 ? da0 ^ dv0 + db1 ^ dw1 ? db0 ^ dw0 ]
t
m m m m m m m = 0
r=1
which is a discretization of (48) integrated over the domain [0;
x]  [0;
t].
Using the complex-valued state variable z = ( ; )T 2 C 2 , @x = , we can rewrite the multi-
symplectic formulation of the nonlinear Schrodinger equation in a more compact notation as
i@t ? @x  = V 0 (j j2 ) ;
@x = 
and the general form of a multi-symplectic Gauss-Legendre discretization can be found in Fig. 9.

m +
x X aij @x j;m ;
s
i;m = 0
j =1
m X s
i;m = 0 +
x aij @xj;m ;
j =1
m = m +
x X bi @x i;m ;
s
1 0
i=1
m X s
1 = 0 +
x bi @x i;m ;
i=1

0 X r
i;m = i +
t a~mn @t i;n ;
n=1
1 0 X~r
i = i +
t bm @t i;m ;
m=1

i@t i;m ? @x i;m = V 0 (j i;m j2 j) i;m ;

@x i;m = i;m :

Figure 9: Multi-Symplectic Gauss-Legendre Collocation Methods for the Nonlinear Schrodinger

Equation

20
In particular, the choice r = s = 1 leads again to a Preissmann scheme
i  1 ? 0  = 1 1=2 ? 1=2  + V 0 (j 1=2 j2 ) 1=2 ;

t  1=2 1=2 
x 1 0 1=2 1=2
1 1=2 1=2 = 1=2

x 1 ? 0 1=2
with
m 1 m m 1=2 = 1 ? 1 + 0
; 1=2 = 1 ?1 + 0
;
1=2 = 2 ( 1 + 0 ) ; i 2 i i i 2 i i
m; i = 1; 2, and
1=2 1 ? 1 1 0 0
1=2 1 ? 1 1 0 0

1=2 = 4 1 + 0 + 1 + 0 ; 1=2 = 4 1 + 0 + 1 + 0 :
We like to point out that any multi-symplectic Gauss-Legendre discretization will exactly preserve
a discrete version of the conservation law (49). This follows from the quadratic form of the conser-
vation law and the fact that quadratic conservation laws are exactly preserved by Gauss-Legendre
methods.
When it comes to the integration of an initial-boundary value (Cauchy) problem, one could
discretize in space using a Gauss-Legendre method and replace the temporal discretization by any
convenient symplectic discretization of the spatially truncated system. Similar to what has been
shown in x4 for the nonlinear wave equation, we would again obtain multi-symplectic methods.

7 Concluding Remarks
Theoretical results indicate [3, 4, 5] that the multi-symplectic structure of certain PDEs is crucial
for a deeper understanding of wave phenomena. One of the nice features is that it is a strictly local
concept and that it can be formulated as a conservation law involving dierential two forms. It is
thus also a very natural generalization of the conservation of symplecticity for canonical Hamiltonian
systems. The results of this paper show that it is possible to nd higher order multi-symplectic
methods in a rather general setting again similar to the nite dimensional situation.
The suggested Gauss-Legendre discretizations for the one dimensional nonlinear wave equation
(1) naturally generalize to any multi-symplectic PDE (2). Furthermore, they can also be applied to
higher-dimensional nonlinear wave equations like, for example,
@tt u ? @xx u ? @yy u = ?V 0 (u)
and, more generally, to multi-symplectic PDEs like, for example,
M@tz + K@xz + L@y z = rz S (z):
This also implies that the schemes can be applied to the corresponding stationary problems charac-
terised by @t z = 0.
Further results on multi-symplectic methods in the context of nite volume methods can be
found in [17].

Acknowledgement. We like to thank Tom Bridges for the introduction to his framework of multi-
symplectic PDEs and for many stimulating discussions during the preparation of this paper.

21
References
[1] M. B. Abbott & D. R. Basco, Computational Fluid Dynamics, Longman Scientic & Tech-
nical, 1989.
[2] G. Benettin & A. Giorgilli, On the Hamiltonian interpolation of near to the identity sym-
plectic mappings with application to symplectic integration algorithms, J. Stat. Phys. 74 (1994)
1117{1143.
[3] Th. J. Bridges, Multi-symplectic structures and wave propagation, Math. Proc. Camb. Phil.
Soc. 121 (1997) 147{190.
[4] Th. J. Bridges, A geometric formulation of the conservation of wave action and its implic-
ations for signature and the classication of instabilities, Proc. R. Soc. Lond. A 453 (1997),
1365{1395.
[5] Th. J. Bridges & G. Derks, Unstable eigenvalues, and the linearization about solitary waves
and fronts with symmetry, Proc. R. Soc. Lond. A (1999), to appear.
[6] Th. J. Bridges & S. Reich, Multi-symplectic integrators: numerical schemes for Hamiltonian
PDEs that conserve symplecticity, technical report.
[7] G. J. Cooper, Stability of Runge-Kutta methods for trajectory problems, IMA J. Numer. Anal.
7 (1987), 1{13.
[8] D. B. Duncan, Symplectic nite dierence approximations of the nonlinear Klein-Gordon
equation, SIAM J. Numer. Anal. 34 (1997), 1742{1760.
[9] B. Fornberg, A Practical Guide to Pseudospectral Methods, Cambridge University Press,
1998.
[10] S. Jimenez, Derivation of the discrete conservation laws for a family of nite dierence
schemes, Applied Mathematics and Computation 64 (1994), 13{45.
[11] J. E. Marsden, G. P. Patrick & S. Shkoller, Multisymplectic geometry, variational
integrators, and nonlinear PDEs, Comm. in Math. Phys., 199 (1999), 351-395.
[12] J. E. Marsden & S. Shkoller, Multisymplectic geometry, covariant Hamiltonians and water
waves, Math. Proc. Camb. Phil. Soc. 125 (1999), 553{575.
[13] R. I. McLachlan, Symplectic integration of Hamiltonian wave equations, Numer. Math. 66
(1994) 465{492.
[14] R. I. McLachlan, On the numerical integration of ordinary dierential equations by symmetric
composition methods, SIAM J. Sci. Comput. 16 (1995), 151{168.
[15] S. Li & L. Vu-Quoc, Finite dierence calculus invariant structure of a class of algorithms for
the nonlinear Klein-Gordon equation, SIAM J. Numer. Anal. 32 (1995), 1839{1875.
[16] S. Reich, Backward error analysis for numerical integrators, SIAM Numer. Anal. (1999) to
appear.
[17] S. Reich, Finite volume methods for multi-symplectic PDEs, technical report, 1999.
[18] J. M. Sanz-Serna & M. P. Calvo, Numerical Hamiltonian Systems, Chapman & Hall,
London, 1994.

22
[19] W. Strauss & L. Vazquez, Numerical solution of a nonlinear Klein-Gordon equation,
J. Comput. Phys. 28 (1978), 271{278.
[20] A. P. Veselov, Integrable discrete-time systems and dierence operators, Functs. Anal.
Prilozhen. 22 (1988), 1{13.
[21] L. Vu-Quoc & S. Li, Invariant-conserving nite dierence algorithms for the nonlinear Klein-
Gordon equation, Computer Methods in Applied Mechanics and Engineering 107 (1993), 341{
391.

Appendix: Energy Conserving Preissmann Scheme

The Preissmann scheme (35) applied to the wave equation (1) results in the system
u11=2 ? u01=2 v11=2 + v10=2

t = 2 ; (50)
v11=2 ? v10=2 w11=2 ? w01=2 ? V 0 (u1=2 ) ;

t =
x 1=2 (51)
u11=2 ? u10=2 = w11=2 + w01=2 : (52)

x 2
Following Li & Vu-Quoc [21, 15], we modify the equation (51) to
v11=2 ? v10=2 w11=2 ? w01=2 ? V (u11=2 ) ? V (u01=2 ) ;

t =
x u11=2 ? u01=2 (53)

We like to show that this modied scheme exactly conserves the discrete energy conservation law
E11=2 ? E10=2 F11=2 ? F01=2

t +
x =0
with the abbreviations
 2  2
E1n=2 = 21 w1n=2 + v1n=2 + V (un1=2 ) and Fm1=2 = ?vm1=2 wm1=2 ;
n; m = 0; 1. This discrete energy conservation law corresponds to the formula (36) with r = s = 1,
b1 = ~b1 = 1, and i = m = 1=2.
For the subsequent derivation it is crucial to observe that the solutions of (50)-(52) as well as of
(50) and (52)-(53) satisfy
w11=2 ? w10=2 1  u11 ? u10 u01 ? u00  1  u11 ? u01 u10 ? u00  v11=2 ? v01=2

t =
t
x ?
x =
x
t ?
x =
x :
We now premultiply (52) by (w11=2 ? w10=2 )=
t and use the above identity to obtain
 1 2  0 2
w1=2 ? w1=2 w11=2 ? w10=2 u11=2 ? u10=2 v11=2 ? v01=2 w11=2 + w01=2
2
t =
t
x =
x 2 : (54)
Next we premultiply (50) by (v11=2 ? v10=2 )=
t which yields
 1 2  0 2
v1=2 ? v1=2 v1 ? v0 u1 ? u0
2
t = 1=2
t 1=2 1=2
t 1=2 :

23
Finally, equation (53) is premultiplied by (u11=2 ? u01=2)=
t which, using the previous equation, results
in  1 2  0 2
v1=2 ? v1=2 v11=2 + v10=2 w11=2 ? w01=2 V (u11=2 ) ? V (u01=2 )
2
t = 2
x ?
t : (55)
Upon combining (54) and (55), we get
E11=2 ? E10=2 v11=2 + v01=2 w11=2 ? w01=2 + v11=2 ? v01=2 w11=2 + w01=2

t = 2
x
x 2
v 1=2 w1=2 ? v1=2 w1=2
= 1 1
x 0 0
1=2 1=2
= ? F1
?xF0
as desired. The system (50) and (52)-(53) does not, in general preserve the momentum conservation
law
I11=2 ? I10=2 G11=2 ? G10=2

t +
t =0
with
1
 2  2
I1=2 = ?v1=2 w1=2 and Gm = 2 vm1=2 + wm1=2 ? V (u1m=2 ) :
n n n 1=2

To obtain a momentum conserving algorithm, (53) has to be replaced by

v11=2 ? v10=2 w11=2 ? w01=2 V (u11=2 ) ? V (u10=2 )
= ? :

t
x u11=2 ? u10=2

24