ARTICLE IN PRESS APNUM:2245

Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.1 (1-19)
Applied Numerical Mathematics ()
Contents lists available at ScienceDirect
Applied Numerical Mathematics
www.elsevier.com/locate/apnum
Exponential RungeKutta methods for the Schrdinger equation
Guillaume Dujardin

Department of Applied Mathematics and Theoretical Physics, Center for Mathematical Sciences, Wilberforce road, Cambridge CB3 0WA, UK
a r t i c l e i n f o a b s t r a c t
Article history:
Received 23 January 2008
Received in revised form 3 February 2009
Accepted 4 February 2009
Available online xxxx
Keywords:
Exponential integrators
RungeKutta methods
Schrdinger equation
We consider exponential RungeKutta methods of collocation type, and use them to solve
linear and semi-linear Schrdinger Cauchy problems on the d-dimensional torus. We prove
that in both cases (linear and non-linear) and with suitable assumptions, s-stage methods
are of order s and we give sucient conditions to achieve orders s + 1 and s + 2.
We show and explain the effects of resonant time steps that occur when solving linear
Schrdinger problems on a nite time interval with such methods. This work is inspired by
[M. Hochbruck, A. Ostermann, Exponential RungeKutta methods for parabolic problems,
Appl. Numer. Math. 53 (24) (2005) 323-339], where exponential RungeKutta methods of
collocation type are applied to parabolic Cauchy problems. We compare our results with
those obtained for parabolic problems and provide numerical experiments for illustration.
2009 IMACS. Published by Elsevier B.V. All rights reserved.
1. Introduction
Exponential integrators are specic geometric integrators. They have become very popular recently for the numerical
integration of rst order in time problems (see for example [15], where their multiple discoveries from the 60s are recalled).
For example, some specic exponential integrators have been used for solving the non-linear Schrdinger equation (see for
example [3] where the numerical analysis is made in one periodical space dimension).
Exponential RungeKutta methods are particular exponential integrators. These methods have been derived and analysed
for semi-linear parabolic Cauchy problems (see for example [14] for collocation methods and [13] for explicit methods).
This article deals with such exponential RungeKutta methods applied to the linear and semi-linear Schrdinger equa-
tions, namely

t
u(t, x) u(t, x) = f (t, x) (linear)
and

t
u(t, x) u(t, x) = f (t, u) (semi-linear),
considered as Cauchy problems in time (no space discretisation is made). Of course, stands for the imaginary unit
(
2
= 1). The analysis in these cases is different from the one performed for parabolic problems since the spectra of
the linear operators are different. We provide a numerical analysis of exponential RungeKutta methods of collocation type
applied to linear and semi-linear problems on the d-dimensional torus (d N

). We show that s-stage collocation methods

are of order s for suitable Sobolev norms. Moreover, we give algebraic sucient conditions on the collocation points of an
s-stage method to achieve orders s +1 and s +2 when solving Schrdinger Cauchy problems. These results need additional
assumptions, for example on the non-linear term of semi-linear problems. These assumptions are fullled in many cases of
*
Tel.: (+44) 1223 760403; fax: (+44) 1223 339378.
E-mail address: g.dujardin@damtp.cam.ac.uk.
0168-9274/$30.00 2009 IMACS. Published by Elsevier B.V. All rights reserved.
doi:10.1016/j.apnum.2009.02.002
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.2 (1-19)
2 G. Dujardin / Applied Numerical Mathematics ()
interest, for example in the case of the cubic non-linear Schrdinger equation. We also mention, quantify and explain the
effects of resonant time steps that appear when solving linear Schrdinger problems on a nite time interval.
Many other results for the nite-time integration of highly oscillatory (discretisations of) PDEs exist in the literature.
For second-order in time problems, like wave equations, we mention the mollied impulse method analysed in [8] and
generalised in [16], the Gautschi-type methods analysed in [12] and the analysis carried out in [9] based on a one-step
formulation.
This paper is concerned with the numerical integration of Hamiltonian PDEs over a nite time interval. For results on
numerical integration of Hamiltonian PDEs over long times, the reader may refer for example to [4,5,10] for results on the
non-linear wave equation obtained via modulated Fourier expansions and to [7] for results on splitting methods applied to
the linear Schrdinger equation with small potential.
Let us mention that many other geometric integrators exist to integrate numerically Schrdinger Cauchy problems. For
example, some of them take advantage of the conservation of phase space properties by considering multisymplectic for-
mulations (see for example [2]).
This article is organised as follows: in Section 2 we deal with exponential RungeKutta methods of collocation type
applied to linear Schrdinger Cauchy problems. We rst introduce some notation for these methods applied to linear prob-
lems. Then we look at order conditions for s-stage methods. We show that these methods are of order at least s and give
sucient conditions to achieve orders s + 1 and s + 2. Moreover, we study the effect of resonant time steps on the order
constants. In Section 3, we look at exponential RungeKutta methods of collocation type applied to semi-linear Schrdinger
Cauchy problems. We set up some notation for these methods applied to non-linear problems. We prove, with suitable as-
sumptions, that s-stage methods provide unique numerical solutions and are of order at least s. Moreover, we give sucient
order conditions for s-stage methods to be of order s + 1 and s + 2. We conclude the article with numerical experiments
and some comments.
2. Linear problems
2.1. Notation
We denote by T the one-dimensional torus R/(2Z). For all given positive integers d, we denote by T
d
the d-dimen-
sional torus. For x = (x
1
, . . . , x
d
) R
d
and y = (y
1
, . . . , y
d
) R
d
, we denote by x. y the real number x
1
y
1
+ +x
d
y
d
and by
|x|
2
the non-negative real number (x
2
1
+ + x
2
d
)
1/2
. For k Z
d
, we denote by p
k
the function on T
d
dened for x T
d
by
p
k
(x) = e
kx
. L
2
(T
d
) (or simply L
2
) denotes the set of (classes of) complex functions f on T
d
such that
_
T
d
| f (x)|
2
dx < +,
endowed with the norm f
L
2
(T
d
)
= ((2)
d
_
T
d
| f (x)|
2
dx)
1/2
. An unbounded linear operator A acting on L
2
is said to be
diagonal if for all k Z
d
, p
k
is in the domain of A and there exists a
k
C such that Ap
k
=
k
p
k
. For example, the Laplace
operator acting on L
2
(T
d
) is diagonal since for all k Z
d
, p
k
= |k|
2
2
p
k
. Another example is the identity operator on
L
2
denoted by Id. For all such A and all functions from C to itself, we denote by (A) the unbounded diagonal linear
operator acting on L
2
whose domain is the set of linear combinations of functions (p
k
)
kZ
d dened for all k Z
d
by
(A)p
k
= (
k
)p
k
. (2.1)
For all functions f L
2
and all k Z
d
, we denote by

f
k
the Fourier coecient (2)
d
_
T
d
f (x)e
kx
dx. For R
+
, we
denote by H

(T
d
) (or simply H

) the space of (classes of) complex functions f L

2
(T
d
) such that

kZ
d
\{0}
|

f
k
|
2
|k|
2
2
<
+, endowed with the norm
f
H
=
_
|

f
0
|
2
+

kZ
d
\{0}
|

f
k
|
2
|k|
2
2
_
1/2
.
Note that L
2
= H
0
with the same norm. With these notations, if is a function from C to C bounded by M 0 on R, then
for all h > 0 and 0, (h) is a bounded linear operator from H

to itself with
1
(h)
H

H
M. In particular,
we have for all 0, e
h

H
= 1.
Following [14], we set for all k N

z C

,
k
(z) =
1
z
k
_
e
z

k1

p=0
z
p
p!
_
, (2.2)
1
Of course, for all 0 and all linear operators A from H

to itself, we denote
A
H

H
= sup
vH

v=0
Av
H

v
H

.
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.3 (1-19)
G. Dujardin / Applied Numerical Mathematics () 3
and
k
(0) =
1
k!
. We also set
0
= exp. Note that for all k N,
k
is holomorphic on C, and is bounded on R. For all
unbounded linear diagonal operator A and all h > 0, we have
k N,
k+1
(hA) =
1
h
k+1
h
_
0
e
(h)A

k
k!
d. (2.3)
In order to compute numerical solutions of the following Cauchy problem
_

t
u(t, x) + A u(t, x) = f (t, x), (t, x) [0, T ] T
d
,
u(0, x) = u
0
(x), x T
d
,
(2.4)
where T > 0, A = , u
0
and f are given, we consider the following numerical methods, called exponential RungeKutta
methods of collocation type; we refer to [14] for a derivation of such methods for semi-linear problems based on the variation-
of-constants formula:
n {0, . . . , N},

t
n
=nh,
u
n+1
= e
hA
u
n
+h
s

i=1
b
i
(hA) f (t
n
+c
i
h),
(2.5)
where h > 0, s N

, (c
1
, . . . , c
s
) [0, 1]
s
are given and N = T /h. We assume that for all (i, j) {1, . . . , s}
2
, c
i
= c
j
if i = j.
The operators b
i
(hA) are dened by
i {1, . . . , n}, b
i
(hA) =
1
h
h
_
0
e
(h)A
l
i
() d, (2.6)
where for all i {1, . . . , s}, l
i
is the i-th Lagrange polynomial with respect to the points (c
j
h)
j{1,...,s}
:
i {1, . . . , s}, l
i
() =
s

j=1, j=i
/h c
j
c
i
c
j
. (2.7)
Note that for all i {1, . . . , s}, b
i
span(
0
, . . . ,
s1
). We therefore derive that b
i
is holomorphic on C and is bounded
on R.
In the case A = 0, the linear PDE (2.4) reduces to the collection of linear ODEs for x T
d
du
dt
(t, x) = f (t, x), t [0, T ],
with initial conditions u(0, x) = u
0
(x). Moreover, the exponential RungeKutta method (2.5) reduces to a classical Runge
Kutta method called the underlying RungeKutta method with coecients (b
i
(0))
i{1,...,s}
since for all i {1, . . . , s}, b
i
(hA) =
b
i
(0)Id.
2.2. Order of exponential RungeKutta methods for linear problems
2.2.1. An s-stage method is of order s
Our rst result shows that an s-stage exponential RungeKutta method of collocation type (2.5) applied to a linear
Schrdinger problem (2.4) is of order at least s, provided that the right-hand side of (2.4) is suciently smooth with
respect to the time, when considered as an L
2
(T
d
)-valued function. This somehow natural result is very similar to the one
obtained in the context of parabolic linear equations in [14] (see Theorem 1). However, note that in the case of the linear
Schrdinger equation, the order constant C does not depend on T .
Theorem 2.1. An s-stage exponential RungeKutta method of collocation type (2.5) applied to a linear problem (2.4) such that
f C
s
([0, T ], L
2
(T
d
)) is of global order s in the sense that there exists a positive constant C depending only on (c
1
, . . . , c
s
) such
that for all h > 0 and all n {0, . . . , N},
_
_
u
n
u(t
n
)
_
_
L
2
Ch
s
t
n
_
0
_
_
f
(s)
()
_
_
L
2
d. (2.8)
Proof. Let us denote for all n {0, . . . , N}, e
n
= u
n
u(t
n
). The variation-of-constants formula for the exact solution u of
problem (2.4) reads for h ]0, T ] and t [0, T h],
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.4 (1-19)
4 G. Dujardin / Applied Numerical Mathematics ()
u(t +h) = e
hA
u(t) +
h
_
0
e
(hs)A
f (t + s) ds. (2.9)
Using Taylor expansions of f in (2.5) and (2.9) and the properties of the Lagrange collocation polynomials (2.7) involved in
the functions (b
i
)
i{1,...,s}
(see (2.6)), one has
e
n+1
= e
hA
e
n
+
h
_
0
e
(h)A

_
0
( )
s1
(s 1)!
f
(s)
(t
n
+) d d

i=1
h
_
0
e
(h)A
l
i
() d
c
i
h
_
0
(c
i
h )
s1
(s 1)!
f
(s)
(t
n
+) d. (2.10)
Recall that e
hA

L
2
L
2 = 1. Denoting
n+1
= e
n+1
e
hA
e
n
, we have that for some constant C depending only on c
1
, . . . , c
s
,

n+1

L
2 Ch
s
t
n+1
_
t
n
_
_
f
(s)
()
_
_
L
2
d.
This inequality and the relation
e
n
=
n1

p=0
e
phA

np
,
complete the proof. 2
Remark 2.2. Notice that for all r > 0, estimate (2.8) holds true with the L
2
(T
d
)-norm replaced by the H
r
(T
d
)-norm provided
that f C
s
([0, T ], H
r
(T
d
)).
2.2.2. Achieving order s +1
If its underlying RungeKutta method is of order s +1, then the exponential RungeKutta method of collocation type (2.5)
applied to the linear Schrdinger problem (2.4) also has order s + 1, provided that we assume that the right-hand side of
(2.4) has higher spatial regularity than just a standard L
2
one (see Theorem 2.4 for a precise statement). This is a difference
with the case of parabolic linear equations studied in [14] (see Theorem 2). Let us rst recall (see [14], formula (12)) that
Lemma 2.3. If the underlying RungeKutta method is of order s +1, then
s

i=1
c
s
i
b
i
(0) =
1
s +1
. (2.11)
We are now able to state the following result:
Theorem 2.4. Assume r 0 is given, the underlying RungeKutta method is of order s +1 and f C
s+1
([0, T ], L
2
(T
d
)) is such that
f
(s)
C
1
([0, T ], H
r+2
). Then, the exponential RungeKutta method of collocation type (2.5) applied to the linear problem (2.4) is of
order s + 1 in the sense that there exists a positive constant C depending only on (c
1
, . . . , c
s
) such that we have for all h > 0 and all
n {0, . . . , N},
_
_
u(t
n
) u
n
_
_
H
r
CTh
s+1
_
_
_
f
(s)
(0)
_
_
H
r+2
+
t
n
_
0
_
_
f
(s+1)
()
_
_
H
r+2
d
_
. (2.12)
Proof. With the notation of the previous proof, using another Taylor series expansion of f in (2.10), we write

n+1
=
(1)
n+1
+
(2)
n+1
,
with

(1)
n+1
=
_
h
_
0
e
(h)A

_
0
( )
s1
(s 1)!
d d
s

i=1
h
_
0
e
(h)A
l
i
() d
c
i
h
_
0
(c
i
h )
s1
(s 1)!
d
_
f
(s)
(t
n
),
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.5 (1-19)
G. Dujardin / Applied Numerical Mathematics () 5
and, after integration by parts,

(2)
n+1
=
h
_
0
e
(h)A

_
0
( )
s
s!
f
(s+1)
(t
n
+) d d
s

i=1
h
_
0
e
(h)A
l
i
() d
c
i
h
_
0
(c
i
h )
s
s!
f
(s+1)
(t
n
+) d. (2.13)
Accordingly, we set
e
(1)
n
=
n1

p=0
e
phA

(1)
np
and e
(2)
n
=
n1

p=0
e
phA

(2)
np
.
Recall that e
h

H
r
H
r = 1. On one hand, for some constant C
1
> 0 depending only on (c
1
, . . . , c
s
),we have
(2)
n+1

H
r
(T
d
)

C
1
h
s+1
_
t
n+1
t
n
f
(s+1)
()
H
r
(T
d
)
d. Therefore, we derive that
_
_
e
(2)
n
_
_
H
r
_
T
d
_
C
1
h
s+1
t
n
_
t
0
_
_
f
(s+1)
()
_
_
H
r
_
T
d
_
d. (2.14)
On the other hand, since

(1)
n+1
=h
s+1
_

s+1
(hA)
1
s!
s

i=1
c
s
i
b
i
(hA)
_
f
(s)
(t
n
),
an Abel summation yields
e
(1)
n
=h
s+1

s+1
(hA)
__
n1

p=0
e
phA
_
f
(s)
(0) +
n2

p=0
_
p

k=0
e
khA
_
_
f
(s)
(t
np2
) f
(s)
(t
np1
)
_
_
,
where
s+1
(z) =
s+1
(z)
1
s!

s
i=1
c
s
i
b
i
(z). Since the underlying RungeKutta method is of order s +1, Lemma 2.3 and the
fact that
s+1
(0) = ((s + 1)!)
1
ensure that
s+1
(0) = 0. Moreover,
s+1
is a holomorphic function on C and is bounded
on R. We derive that there exists a holomorphic function

s+1
that is bounded on R such that z C,
s+1
(z) = z

s+1
(z).
Hence, there exists a positive constant C
2
depending only on (c
1
, . . . , c
s
) such that

s+1
(hA)
H
r
(T
d
)H
r
(T
d
)
C
2
. We
derive that
_
_
e
(1)
n
_
_
H
r
(T
d
)
C
2
h
s+1
__
_
_
_
_
hA
_
n1

p=0
e
phA
_
f
(s)
(0)
_
_
_
_
_
H
r
_
T
d
_
+
n2

p=0
_
_
_
_
_
hA
_
p

k=0
e
khA
_ t
np1
_
t
np2
f
(s+1)
() d
_
_
_
_
_
H
r
(T
d
)
_
.
For all k Z
d
and p {0, . . . , N 1},

h|k|
2
2
p

=0
e
h|k|
2
2

(p +1)h|k|
2
2
T |k|
2
2
. (2.15)
Therefore, for all p {0, . . . , N 1} and u H
r+2
,
_
_
_
_
_
hA
p

=0
e
hA
u
_
_
_
_
_
H
r
T u
H
r+2 . (2.16)
This allows us to bound the H
r
-norm of e
(1)
n
. Together with estimate (2.14), this ensures that inequality (2.12) holds true. 2
Note that the order constant in the previous result depends on T . This is also the case when the exponential RungeKutta
method (2.5) is applied to a linear parabolic problem (see [14], Theorem 2). In the case of linear Schrdinger problems (2.4),
if the underlying RungeKutta method is of order s + 1, it is possible to get an order s + 1 constant that does not depend
on T , with a restriction on the values of the time step. The latter has to be non-resonant in the following sense:
Lemma 2.5. Assume the following non-resonance condition on the time step h > 0: there exists > 0 and > 1 such that
N N

1 e
hN
h

. (2.17)
Then we have for all n {0, . . . , N}, for all r 0 and all u H
r+2+2
(T
d
),
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.6 (1-19)
6 G. Dujardin / Applied Numerical Mathematics ()
_
_
_
_
_
hA
n1

p=0
e
phA
u
_
_
_
_
_
H
r
(T
d
)

u
H
r+2+2
(T
d
)
. (2.18)
Proof. With the help of (2.17), we get for k Z
d
\ {0},

h|k|
2
2
n1

p=0
e
ph|k|
2
2

2h|k|
2
2
1 e
h|k|
2
2

|k|
2+2
2
.
Therefore, estimate (2.18) holds true. 2
Note that the set of time steps h (0, h
0
) that do not satisfy (2.17) has a Lebesgue measure in o(h
0
) as h
0
tends to 0.
See for example [11], Chapter 10, Lemma 6.3.
In addition to the restriction on the values of the time step, a higher spatial regularity of the right-hand side f of (2.4)
is assumed to prove the following:
Theorem 2.6. Assume that r 0, > 0 and > 1 are given. Assume that the underlying RungeKutta method is of order s + 1 and
f C
s+1
([0, T ], L
2
(T
d
)) is such that f
(s)
C
1
([0, T ], H
r+2+2
). Then the exponential RungeKutta method of collocation type (2.5)
applied to the linear problem (2.4) is of order s + 1 for non-resonant time steps in the sense that there exists a positive constant C
depending only on (c
1
, . . . , c
s
) and such that if h > 0 satises (2.17), then we have for all n {0, . . . , N},
_
_
u(t
n
) u
n
_
_
H
r
Ch
s+1
_
_
_
f
(s)
(0)
_
_
H
r+2+2
+
t
n
_
0
_
_
f
(s+1)
()
_
_
H
r+2+2
d
_
. (2.19)
Proof. With the help of Lemma 2.5, estimate (2.16) can be replaced with estimate (2.18) in the proof of Theorem 2.4 to get
_
_
e
(1)
n
_
_
H
r
2
C
2

h
s+1
_
_
_
f
(s)
(0)
_
_
H
r+2+2
+
t
n1
_
t
0
_
_
f
(s+1)
()
_
_
H
r+2+2
d
_
. (2.20)
In addition to (2.14), we derive that (2.19) holds true. 2
The effect of resonant time steps on the global error over the nite time interval [0, T ] is illustrated by numerical
experiments in Section 4. In particular, one can see the difference between time steps satisfying (2.17) for which estimate
(2.19) is sharp and those that do not satisfy (2.17) for which estimate (2.12) is sharp.
2.2.3. Concerning order s +2
As in the case of exponential RungeKutta methods applied to linear parabolic problems studied in [14] (see Theorem 3),
if its underlying RungeKutta method is of order s +2 and one has sucient spatial regularity in the right-hand side of (2.4),
then the exponential RungeKutta method (2.5) applied to the linear problem (2.4) also has order s +2. Before giving precise
statements, let us recall the following: (see [14], Lemma 3)
Lemma 2.7. If the underlying RungeKutta method is of order s +2, then
s

i=1
c
s+1
i
b
i
(0) =
1
s +2
, (2.21)
s

i=1
c
s
i
b

i
(0) =
1
(s +1)(s +2)
. (2.22)
We are now able to prove the following:
Theorem 2.8. Assume that r 0 is given, the underlying RungeKutta method is of order s +2 and f C
s+2
([0, T ], L
2
(T
d
)) is such
that f
(s)
C
2
([0, T ], H
r+4
). Then the exponential RungeKutta method of collocation type (2.5) applied to the linear Schrdinger
equation (2.4) is of order s +2 in the sense that there exists a positive constant C depending only on (c
1
, . . . , c
s
) such that for all h > 0
and all n {0, . . . , N}, we have
_
_
u(t
n
) u
n
_
_
H
r
CTh
s+2
_
_
_
f
(s)
(0)
_
_
H
r+4
+
_
_
f
(s+1)
(0)
_
_
H
r+2
+
t
n
_
0
_
_
f
(s+1)
()
_
_
H
r+4
d +
t
n
_
0
_
_
f
(s+2)
()
_
_
H
r+2
d
_
.
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.7 (1-19)
G. Dujardin / Applied Numerical Mathematics () 7
Proof. Let
(2)
n+1
be dened by (2.13) and write
(2)
n+1
=
(2,1)
n+1
+
(2,2)
n+1
with

(2,1)
n+1
=h
s+2
_

s+2
(hA)
1
(s +1)!
s

i=1
c
s+1
i
b
i
(hA)
_
f
(s+1)
(t
n
),
and

(2,2)
n+1
=
h
_
0
e
(h)A

_
0
( )
s+1
(s +1)!
f
(s+2)
(t
n
+) d d
s

i=1
h
_
0
e
(h)A
l
i
() d
c
i
h
_
0
(c
i
h )
s+1
(s +1)!
f
(s+2)
(t
n
+) d.
One deduces that there exists C > 0 depending only on (c
1
, . . . , c
s
) such that
_
_

(2,2)
n+1
_
_
H
r
Ch
s+2
t
n+1
_
t
n
_
_
f
(s+2)
()
_
_
H
r
d.
Therefore,
_
_
_
_
_
n1

p=0
e
phA

(2,2)
np
_
_
_
_
_
H
r
Ch
s+2
n1

p=0
t
np
_
t
np1
_
_
f
(s+2)
()
_
_
H
r
d Ch
s+2
t
n
_
t
0
_
_
f
(s+2)
()
_
_
H
r
d. (2.23)
Since the underlying RungeKutta method is of order s + 2, Lemma 2.7 ensures that

s
i=1
c
s+1
i
b
i
(0) =
1
s+2
(see (2.21)).
Moreover,
s+2
(0) = ((s +2)!)
1
. Therefore, there exists a holomorphic function
s+2
on C such that
z C,
1
(s +1)!
s

i=1
c
s+1
i
b
i
(z)
s+2
(z) = z
s+2
(z).
Moreover,
s+2
is bounded on R. Hence, we have

(2,1)
n+1
= hAh
s+2

s+2
(hA) f
(s+1)
(t
n
),
and by an Abel summation,
n1

p=0
e
phA

(2,1)
np
=h
s+2

s+2
(hA)
_
hA
_
n1

p=0
e
phA
_
f
(s+1)
(0)
+
n2

p=0
hA
_
p

k=0
e
khA
_
_
f
(s+1)
(t
np2
) f
(s+1)
(t
np1
)
_
_
.
We derive with estimate (2.15) that there exists a positive constant C
1
depending only on (c
1
, . . . , c
s
) such that
_
_
_
_
_
n1

p=0
e
phA

(2,1)
np
_
_
_
_
_
H
r
C
1
Th
s+2
_
_
_
f
(s+1)
(0)
_
_
H
r+2
+
t
n1
_
t
0
_
_
f
(s+2)
()
_
_
H
r+2
d
_
. (2.24)
To complete the proof, since the underlying RungeKutta method is of order s + 2, we get by Lemmas 2.3 and 2.7 that

s
i=1
c
s
i
b
i
(0) =
1
s+1
and

s
i=1
c
s
i
b

i
(0) =
1
(s+1)(s+2)
(see relation (2.22)). Therefore, there exists a holomorphic function
s+2
on C such that
s+1
(z)
1
s!

s
i=1
c
s
i
b
i
(z) = z
2

s+2
(z). Moreover,
s+2
is bounded on R. As before, we get
n1

p=0
e
phA

(1)
np
= hh
s+1

s+2
(hA)
_
hA
_
n1

p=0
e
phA
_
Af
(s)
(0)
+
n2

p=0
hA
_
p

k=0
e
khA
_
A
_
f
(s)
(t
np2
) f
(s)
(t
np1
)
_
_
.
Therefore, using inequality (2.15), there exists a positive constant C
2
depending only on (c
1
, . . . , c
2
) such that
_
_
_
_
_
n1

p=0
e
phA

(1)
np
_
_
_
_
_
H
r
C
2
Th
s+2
_
_
_
f
(s)
(0)
_
_
H
r+4
+
t
n1
_
t
0
_
_
f
(s+1)
()
_
_
H
r+4
d
_
. (2.25)
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.8 (1-19)
8 G. Dujardin / Applied Numerical Mathematics ()
The conclusion follows by adding (2.23), (2.24) and (2.25) together. 2
In the previous theorem, the order constant depends on T . As in the previous section (see Theorem 2.6), if the underlying
RungeKutta method is of order s + 2, then the exponential RungeKutta method (2.5) applied to the linear Schrdinger
problem (2.4) is of order s + 2 with an order constant independent of T provided that the time step is non-resonant and
the right-hand side f of (2.4) has higher spatial regularity. A precise statement is the following:
Theorem 2.9. Assume that r 0, > 0 and > 1 are given as in Theorem 2.6. Assume that the underlying RungeKutta method
is of order s + 2 and that f C
s+2
([0, T ], L
2
(T
d
)) is such that f
(s)
C
2
([0, T ], H
r+2+4
(T
d
)). Then the exponential RungeKutta
method of collocation type (2.5) applied to the linear Schrdinger equation (2.4) is of order s + 2 for non-resonant time steps in the
sense that there exists a positive constant C depending only on (c
1
, . . . , c
s
) and such that if h > 0 satises (2.17), then we have for
all n {0, . . . , N},
_
_
u(t
n
) u
n
_
_
H
r
Ch
s+2
_
_
_
f
(s)
(0)
_
_
H
r+2+4
+
_
_
f
(s+1)
(0)
_
_
H
r+2+2
+
t
n
_
0
_
_
f
(s+1)
()
_
_
H
r+2+4
d +
t
n
_
0
_
_
f
(s+2)
()
_
_
H
r+2+2
d
_
.
Proof. With the help of Lemma 2.5, estimate (2.24) can be replaced by the following estimate
_
_
_
_
_
n1

p=0
e
phA

(2,1)
np
_
_
_
_
_
H
r

C
1

h
s+2
_
_
_
f
(s+1)
(0)
_
_
H
r+2+2
+
t
n1
_
t
0
_
_
f
(s+2)
()
_
_
H
r+2+2
d
_
, (2.26)
where C
1
> 0 depends only on (c
1
, . . . , c
n
). The same lemma yields the existence of a positive constant C
2
depending only
on (c
1
, . . . , c
s
) such that
_
_
_
_
_
n1

p=0
e
phA

(1)
np
_
_
_
_
_
H
r

C
2

h
s+2
_
_
_
f
(s)
(0)
_
_
H
r+2+4
+
t
n1
_
t
0
_
_
f
(s+1)
()
_
_
H
r+2+4
d
_
. (2.27)
This estimate replaces inequality (2.25). As in the proof of the previous theorem, the conclusion follows by adding (2.23),
(2.26) and (2.27) together. 2
3. Semi-linear problems
3.1. Notation
We consider the following semi-linear Cauchy problem
_

t
u(t, x) + Au(t, x) = g
_
t, u(t, x)
_
, (t, x) [0, T ] T
d
,
u(0, x) = u
0
(x), x T
d
(3.1)
where T > 0, A = , r 0, u
0
H
r
(T
d
) and g are given. In the following and in many applications, for all t [0, T ],
g(t) comes from a non-linear function from C to C. For example, for the so-called cubic non-linear Schrdinger equation,
g(t, u) = g(t)(u) = i|u|
2
u. In order to solve numerically the problem (3.1), we consider the following numerical methods,
also called exponential RungeKutta methods of collocation type. We also refer to [14] for a derivation of such methods for
semi-linear problems based on the variation-of-constants formula. We assume that s N

, c
1
, . . . , c
s
[0, 1], h > 0 are
given and set N = T /h. We also assume that if (i, j) {1, . . . , s}
2
are such that i = j, then c
i
= c
j
. We denote for all
n {0, . . . , N}, t
n
=nh. For such n, if u
n
H
r
(T
d
) is a given approximation of the exact solution u(t
n
) of the problem (3.1),
then we construct an approximation u
n+1
of u(t
n+1
) by rst solving the non-linear system consisting of the s following
equations
u
n,i
= e
c
i
hA
u
n
+h
s

j=1
a
i, j
(hA)g(t
n
+c
j
h, u
n, j
), i {1, . . . , s}. (3.2)
The s unknown quantities are u
n,1
, . . . , u
n,s
H
r
(T
d
) and the s
2
coecients (a
i, j
(hA))
(i, j){1,...,s}
2 are dened by
a
i, j
(hA) =
1
h
c
i
h
_
0
e
(c
i
h)A
l
j
() d.
We recall that the (l
j
)
j{1,...,s}
are the Lagrange polynomials dened in (2.7). Then, we dene
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.9 (1-19)
G. Dujardin / Applied Numerical Mathematics () 9
u
n+1
= e
hA
u
n
+h
s

i=1
b
i
(hA)g(t
n
+c
i
h, u
n,i
). (3.3)
Recall that the coecients (b
i
)
i{1,...,s}
are dened in (2.6). Note that for all (i, j) {1, . . . , s}
2
, a
i, j
(hA) span(
0
(c
i
hA),
. . . ,
s1
(c
i
hA)) and the coecients in each linear combination depend only on the choice of the points (c
1
, . . . , c
s
). Hence,
each coecient a
i, j
(z) is bounded on R by a constant depending only on the choice of (c
1
, . . . , c
s
). In the following, C de-
notes a bound of the functions (b
i
(z))
i{1,...,s}
and (a
i, j
(z))
(i, j){1,...,s}
2 on R depending only on the choice of (c
1
, . . . , c
s
).
In the case A = 0, the non-linear Cauchy problem (3.1) reduces to the collection of non-linear ODEs for x T
d
du
dt
(t, x) = g
_
t, u(t, x)
_
, t [0, T ],
with initial values u(0, x) = u
0
(x). Moreover, the exponential RungeKutta method of collocation type (3.2), (3.3) reduces
to a classical RungeKutta method whose coecients are (a
i, j
(0))
(i, j){1,...,s}
2 and (b
i
(0))
i{1,...,s}
, since in that case for all
(i, j) {1, . . . , s}
2
,
a
i, j
(hA) =a
i, j
(0) Id and b
i
(hA) = b
i
(0) Id.
This method is called the underlying RungeKutta method.
For t [0, T ], we denote f (t) = g(t, u(t)), the right-hand side of (3.1), where u denotes the exact solution of the problem.
The space H
r
(T
d
)
s
is a Hilbert space when endowed with the norm
_
_
(u
1
, . . . , u
s
)
_
_
r,s
=
_
s

i=1
u
i

2
H
r
(T
d
)
_
1/2
. (3.4)
For R 0, we denote by B(0, R) the set of u H
r
(T
d
)
s
such that u
r,s
R.
3.2. Existence and uniqueness of the numerical solution
This section is devoted to the proof of existence and uniqueness of the numerical solution (u
0
, . . . , u
N
) provided by the
exponential RungeKutta method (3.2), (3.3) of the semi-linear problem (3.1) for h suciently small and with some suitable
assumptions on the exact solution and the right-hand side of (3.1). We will also give a rst bound for the numerical error.
Our assumptions on the exact solution u of (3.1) and the non-linearity g are the following:
Hypothesis 3.1. The function g satises:
t [0, T ], g(t, 0) = 0.
Hypothesis 3.2. The non-linearity g in the right-hand side of (3.1) is Lipschitz-continuous in the sense that there exists
L > 0 and > 0 such that for all t [0, T ] and for all u, v H
r
(T
d
) satisfying u
H
r
(T
d
)
and v
H
r
(T
d
)
, we have
_
_
g(t, u) g(t, v)
_
_
H
r
(T
d
)
Lu v
H
r
(T
d
)
. (3.5)
Hypothesis 3.3. Eq. (3.1) admits an exact solution u : [0, T ] H
r
(T
d
) that is suciently smooth. In particular, there exists
R > 0 such that for all t [0, T ], u(t)
H
r
(T
d
)
R.
Hypothesis 3.4. The mapping f : [0, T ] H
r
(T
d
) is suciently smooth.
In applications, Hypothesis 3.4 will often be a consequence of Hypothesis 3.3 and of the regularity of the function g.
Note that these assumptions are fullled in many cases of interest, for example in the case of the cubic non-linear
Schrdinger equation, at least when r > d/2 and r d/2 / N. This can be seen, for example, by adapting classical results
of non-linear analysis (see for example [1], Chapter II, Proposition 2.2). Precise statements of the adaptations can be found
in [6].
Under these assumptions, if is suciently big (see (3.7)) and h is suciently small (see (3.6)), then for all n {0, . . . , N}
the non-linear system (3.2) has a unique solution (u
n,1
, . . . , u
n,s
) H
r
(T)
s
such that (u
n,1
, . . . , u
n,s
)
r,s
3R

s. Moreover,
the approximation u
n+1
dened by (3.3) is such that u
n+1

H
s
(T
d
)
2R. To prove these assertions, we start with the
following lemma:
Lemma 3.5. Assume that h
0
> 0 is such that
h
0

_
3CLs
2
_
1
, (3.6)
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.10 (1-19)
10 G. Dujardin / Applied Numerical Mathematics ()
and
3R

s. (3.7)
Then for all h (0, h
0
), t [0, T h] and y H
r
(T
d
) such that y
H
r
(T
d
)
2R, the non-linear system
v
1
= e
c
1
hA
y +h
s

j=1
a
1, j
(hA)g(t +c
j
h, v
j
),
.
.
.
v
s
= e
c
s
hA
y +h
s

j=1
a
s, j
(hA)g(t +c
j
h, v
j
),
admits a unique solution v = (v
1
, . . . , v
s
) B(0, 3R

s).
Proof. For h (0, h
0
), t [0, T h], y H
r
(T
d
) such that y
H
r
(T
d
)
2R, we dene the function
f
h,t,y
:
_
H
r
(T
d
)
s
H
r
(T
d
)
s
(v
i
)
1is
(e
c
i
hA
y +h

s
j=1
a
i, j
(hA)g(t +c
j
h, v
j
))
1is
_
.
Hypothesis 3.2 and inequality (3.7) ensure that for all (v
1
, . . . , v
s
) B(0, 3R

s), we have
_
_
f
t,h,y
(v
1
, . . . , v
s
)
_
e
c
1
hA
y, . . . , e
c
s
hA
y
__
_
r,s
Chs
3/2
L(3R

s).
Since (e
c
1
hA
y, . . . , e
c
s
hA
y)
r,s
2R

s, the triangle inequality and (3.6) imply that

_
_
f
t,h,y
(v
1
, . . . , v
s
)
_
_
r,s
3R

s.
Therefore, f
t,h,y
(B(0, 3R

s)) B(0, 3R

s).
On the other hand, for v = (v
1
, . . . , v
s
) B(0, 3R

s) and w = (w
1
, . . . , w
s
) B(0, 3R

s), by Hypothesis 3.2 we have

_
_
f
t,h,y
(v) f
t,h,y
(w)
_
_
r,s
CLhsv w
r,s
.
Hence, using (3.6), one derives that
_
_
f
t,h,y
(v) f
t,h,y
(w)
_
_
r,s

h
h
0
v w
r,s
.
Since hh
1
0
< 1, equation x = f
t,h,y
(x) has exactly one solution in B(0, 3R

s) by the classical Picard xed-point theorem. 2

We are now able to prove that if h
0
satises another smallness condition, then, for all h (0, h
0
), the exponen-
tial RungeKutta method of collocation type (3.2), (3.3) provides a unique numerical solution with for all n {0, . . . , N},
(u
n,1
, . . . , u
n,s
)
r,s
B(0, 3R

s) and u
n

H
r
(T
d
)
2R. Moreover, such an s-stage method is of order s when applied to the
semi-linear Schrdinger problem (3.1).
Theorem 3.6. Assume Hypotheses 3.13.3 and 3.4 are satised with (3.7). Set
s
= (s + 1)(
1
s!
+
Cs
(s1)!
). Assume that h
0
> 0 satis-
es (3.6), and
h
s
0
R
_

s
e
2CsT L
T
_
0
_
_
f
(s)
()
_
_
H
r
(T
d
)
d
_
1
. (3.8)
Then, for all h (0, h
0
), the exponential RungeKutta method (3.2), (3.3) provides a unique numerical solution such that for all
n {0, . . . , N}, u
n

H
s
(T
d
)
2R and (u
n,1
, . . . , u
n,s
)
r,s
3R

s. Moreover, for all n {0, . . . , N},

_
_
u
n
u(t
n
)
_
_
H
r
(T
d
)

s
e
2CsLT
h
s
t
n
_
0
_
_
f
(s)
()
_
_
H
r
(T
d
)
d. (3.9)
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.11 (1-19)
G. Dujardin / Applied Numerical Mathematics () 11
We recall that for all n {0, . . . , N}, 0 t
n
T .
Proof. Let h (0, h
0
). We prove the result by induction. Assume n {0, . . . , N 1} is such that u
n

H
r
(T
d
)
2R. Lemma 3.5
ensures that the non-linear system (3.2) has a unique solution (u
n,1
, . . . , u
n,s
) B(0, 3R

s). Moreover, Hypotheses 3.3 and

3.4 ensure that, for the exact solution, we have
u(t
n+1
) = e
hA
u(t
n
) +h
s

i=1
b
i
(hA) f (t
n
+c
i
h) +
n+1
, (3.10)
with

n+1
=
h
_
0
e
(h)A

_
0
( )
s1
(s 1)!
f
(s)
(t
n
+) d d h
s

i=1
b
i
(hA)
c
i
h
_
0
(c
i
h )
s1
(s 1)!
f
(s)
(t
n
+) d,
using the quadrature rule properties. Similarly, for all i {1, . . . , s}, we have
u(t
n
+c
i
h) = e
c
i
hA
u(t
n
) +h
s

j=1
a
i, j
(hA) f (t
n
+c
j
h) +
n,i
, (3.11)
where

n,i
=
c
i
h
_
0
e
(c
i
h)A

_
0
( )
s1
(s 1)!
f
(s)
(t
n
+) d d h
s

j=1
a
i, j
(hA)
c
j
h
_
0
(c
j
h )
s1
(s 1)!
f
(s)
(t
n
+) d. (3.12)
Hence, if we denote e
n
= u
n
u(t
n
) and e
n,i
= u
n,i
u(t
n
+c
i
h), then we have
e
n,i
= e
c
i
hA
e
n
+h
s

j=1
a
i, j
(hA)
_
g(t
n
+c
j
h, u
n, j
) f (t
n
+c
j
h)
_

n,i
. (3.13)
For all i {1, . . . , s}, we have e
c
i
hA

H
r
H
r = 1 and hence
e
n,i

H
r
(T
d
)
e
n

H
r
(T
d
)
+hCL
_
s

j=1
e
n, j

H
r
(T
d
)
_
+
n,i

H
r
(T
d
)
. (3.14)
Summing these s inequalities and using (3.6), we get
s

j=1
e
n, j

H
r
(T
d
)
2se
n

H
r
(T
d
)
+2
s

i=1

n,i

H
r
(T
d
)
. (3.15)
On the other hand, subtracting (3.10) from (3.3) yields
e
n+1
= e
hA
e
n
+h
s

i=1
b
i
(hA)
_
g(t
n
+c
i
h, u
n,i
) f (t
n
+c
i
h)
_

n+1
. (3.16)
Hence,
e
n+1

H
r
(T
d
)
e
n

H
r
(T
d
)
+hCL
s

i=1
e
n,i

H
r
(T
d
)
+
n+1

H
r
(T
d
)
. (3.17)
Using (3.15), we thus get
e
n+1

H
r
(T
d
)
e
n

H
r
(T
d
)
+2hCL
_
se
n

H
r
(T
d
)
+
s

j=1

n, j

H
r
(T
d
)
_
+
n+1

H
r
(T
d
)
(1 +2hCsL)e
n

H
r
(T
d
)
+
s

j=1

n, j

H
r
(T
d
)
+
n+1

H
r
(T
d
)

p=0
(1 +2hCsL)
p
_
s

j=1

np, j

H
r
(T
d
)
+
n+1p

H
r
(T
d
)
_
e
2NhCsL
n

p=0
_
s

j=1

np, j

H
r
(T
d
)
+
n+1p

H
r
(T
d
)
_
,
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.12 (1-19)
12 G. Dujardin / Applied Numerical Mathematics ()
using also (3.6). Using relation (2.10) (which still holds for non-linear problems) and relation (3.12), we get that for all
j {1, . . . , s} and all p {0, . . . , n},

np, j

H
r
(T
d
)
,
n+1p

H
r
(T
d
)

_
1
s!
+
Cs
(s 1)!
_
h
s
t
np+1
_
t
np
_
_
f
(s)
()
_
_
H
r
(T
d
)
d. (3.18)
We derive
e
n+1

H
r
(T
d
)
e
2NhCsL
(s +1)
_
1
s!
+
Cs
(s 1)!
_
h
s
t
n+1
_
t
0
_
_
f
(s)
()
_
_
H
r
(T
d
)
d.
In view of (3.8), we derive that e
n+1

H
r
(T
d
)
R. Hence, by triangle inequality, we get u
n+1

H
r
(T
d
)
u(t
n+1
)
H
r
(T
d
)
+
e
n+1

H
r
(T
d
)
R + R and estimate (3.9) holds true. 2
3.3. A sucient condition for order s +1
As in the linear case (see Theorems 2.4 and 2.6), the exponential RungeKutta method of collocation type (3.2), (3.3)
applied to the non-linear Schrdinger equation (3.1) is of order s + 1 provided that the underlying RungeKutta method is
of order s + 1. A precise result is stated in Theorem 3.8. Note that this result is very similar to the one obtained in [14] in
the case of semi-linear parabolic problems (see [14], Theorem 5). Let us start with a discrete Gronwall Lemma:
Lemma 3.7. Let X be a subset
2
of R
+
. Assume that T > 0 and a 0 are given. There exists a positive constant C such that for all
non-negative functions b dened on X and all sequences (
n
)
nN
of non-negative real numbers satisfying for all h X and n N with
0 nh T ,
n
ah

n1
k=0

k
+b(h), we have

n
Cb(h).
Theorem 3.8. Under the hypotheses of Theorem 3.6, if f
(s)
C
1
([0, T ], H
r+2
) and the underlying RungeKutta method is of order
s +1, then there exists a positive constant C > 0 depending only on (c
1
, . . . , c
s
), L, and T such that for all h (0, h
0
), we have for all
n {0, . . . , N},
_
_
u
n
u(t
n
)
_
_
H
r
Ch
s+1
_
_
_
f
(s)
(0)
_
_
H
r+2
+
T
_
0
_
_
f
(s+1)
()
_
_
H
r+2
d +
T
_
0
_
_
f
(s)
()
_
_
H
r
d
_
.
Proof. By induction, relation (3.16) becomes
e
n
=h
_
n1

k=0
e
khA
d
nk1
_

n1

k=0
e
khA

nk
,
where d
k
=

s
i=1
b
i
(hA)(g(t
k
+ c
i
h, u
k,i
) f (t
k
+ c
i
h)). As in the proof of Theorem 2.4, there exists a constant C
1
> 0
depending only on (c
1
, . . . , c
s
) such that
_
_
_
_
_
n1

k=0
e
khA

nk
_
_
_
_
_
H
r
C
1
h
s+1
_
_
_
f
(s)
(0)
_
_
H
r+2
+
t
n1
_
0
_
_
f
(s+1)
()
_
_
H
r+2
d
_
. (3.19)
On the other hand, with Hypothesis 3.2, we have
_
_
_
_
_
n1

k=0
e
khA
d
nk1
_
_
_
_
_
H
r
CL
n1

k=0
s

i=1
e
k,i

H
r .
Then, using (3.15), we deduce
_
_
_
_
_
n1

k=0
e
khA
d
nk1
_
_
_
_
_
H
r
2sCL
n1

k=0
e
k

H
r +2CL
n1

k=0
s

i=1

k,i

H
r .
Estimate (3.18) provides the existence of a constant C
2
> 0 depending only on (c
1
, . . . , c
s
) and s such that
2
Namely, X will be (0, h
0
) or the set of h (0, h
0
) satisfying (2.17).
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.13 (1-19)
G. Dujardin / Applied Numerical Mathematics () 13
n1

k=0
s

i=1

k,i

H
r C
2
h
s
t
n
_
0
_
_
f
(s)
()
_
_
H
r
d.
Hence,
e
n

H
r 2shCL
n1

k=0
e
k

H
r +h
s+1
_
2CLC
2
T
_
0
_
_
f
(s)
()
_
_
H
r
d + C
1
_
_
_
f
(s)
(0)
_
_
H
r+2
+
T
_
0
_
_
f
(s+1)
()
_
_
H
r+2
d
__
.
We conclude with the help of Lemma 3.7 with
n
= e
n

H
r , a = 2sCL and b equal the term multiplied by h
s+1
in the
previous estimate. 2
Remark 3.9. Note that in this non-linear case, the order constant C appearing in Theorem 3.8 depends a priori exponentially
on T . This contrasts the linear case (see Theorems 2.4 and 2.6).
Remark 3.10. We could write a counterpart of Theorem 3.8 for non-resonant time steps with suitable assumptions on f
essentially by modifying inequality (3.19). However, we would still get an order constant depending on T (see Theorem 2.6
for the corresponding result for linear problems) and the fact is that we did not manage to observe resonances for non-linear
problems (see Section 4 for numerical experiments).
3.4. Achieving order s +2
In order to give sucient algebraic conditions on the coecients of the underlying RungeKutta method for the s-stage
exponential RungeKutta method of collocation type (3.2), (3.3) to be of order s + 2, we reinforce our hypothesis on the
smoothness of the non-linearity g. The function g is assumed to be smooth in the sense of functions from R
2
to R
2
. For
convenience, we denote, for example, for v C=R
2
,
g
u
(t, v) the R-linear mapping from C=R
2
to C=R
2
corresponding
to the rst derivative of g(t, u) as a function of u R
2
. For v H
r
(T
d
), we also denote by
g
u
(t, v) the induced mapping
between functions on T
d
dened for a function w on T
d
by
x T
d
,
g
u
(t, v)(w)(x) =
g
u
_
t, v(x)
__
w(x)
_
.
Our additional hypotheses on the non-linearity g are the following:
Hypothesis 3.11. Assume that there exists positive constants L
1
and

L
1
such that
t [0, T ], v H
r
_
T
d
_
,
_
_
_
_
g
u
_
t, u(t)
_
v
_
_
_
_
H
r
(T
d
)
L
1
v
H
r
(T
d
)
,
and
t [0, T ], v H
r+2
_
T
d
_
,
_
_
_
_
g
u
_
t, u(t)
_
v
_
_
_
_
H
r+2
(T
d
)

L
1
v
H
r+2
(T
d
)
.
Hypothesis 3.12. Assume that there exists a positive constant L
2
such that for all t [0, T ] and u, v, w H
r
(T
d
) satisfying
max{u
H
r , v
H
r , w
H
r } 2R, one has
_
_
_
_
g(t, v) g(t, w)
g
u
(t, u)(v w)
_
_
_
_
H
r
L
2
_
u v
H
r + u w
H
r
_
v w
H
r .
Hypothesis 3.13. Assume that there exists positive constants L
3
and

L
3
such that for all t [0, T ] and v, w H
r+2
(T
d
), one
has
_
_
_
_

2
g
u
2
_
t, u(t)
_
(v, w)
_
_
_
_
H
r
L
3
v
H
r w
H
r ,
_
_
_
_

2
g
tu
_
t, u(t)
_
v
_
_
_
_
H
r
L
3
v
H
r ,
_
_
_
_

2
g
u
2
_
t, u(t)
_
(v, w)
_
_
_
_
H
r+2

L
3
v
H
r+2 w
H
r+2 ,
and
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.14 (1-19)
14 G. Dujardin / Applied Numerical Mathematics ()
_
_
_
_

2
g
tu
_
t, u(t)
_
v
_
_
_
_
H
r+2

L
3
v
H
r+2 .
Such hypotheses on g are coupled with (and may imply) the following one on the exact solution u of problem (3.1) that
reinforces Hypothesis 3.3:
Hypothesis 3.14. Eq. (3.1) has an exact solution u : [0, T ] H
r+2
(T
d
) that is suciently smooth. In particular, u
0
H
r+2
(T
d
).
Notice that Hypotheses 3.11, 3.12 and 3.13 are satised in many cases of interest, for example for the cubic Schrdinger
equation when r >d/2 with r d/2 / N provided that u is suciently smooth (see [6] for details).
Under these hypotheses, we prove that the exponential RungeKutta method (3.2), (3.3) applied to the semi-linear
Schrdinger equation (3.1) is of order s +2 provided that the underlying RungeKutta method is of order s +2.
Theorem 3.15. Assume that r 0 is given. Assume that the underlying RungeKutta method is of order s +2, and that the hypotheses
of Theorem 3.6, Hypotheses 3.113.14 are fullled. If f C
s+2
([0, T ], H
r+4
), then there exists a positive constant C > 0 such that for
all h (0, h
0
) and all n {0, . . . , N}, one has
_
_
u(t
n
) u
n
_
_
H
r
Ch
s+2
.
Remark 3.16. Of course, the constant C appearing in Theorem 3.15 also involves the nal time T > 0 and integrals of norms
of the time-derivatives of f over [0, T ].
Proof. Let us set for n {0, . . . , N}
J
n
=
g
u
_
t
n
, u(t
n
)
_
,
and

d
n,i
= g(t
n
+c
i
h, u
n,i
) g
_
t
n
+c
i
h, u(t
n
+c
i
h)
_
J
n
e
n,i
.
As before, subtracting (3.10) from (3.3) yields
e
n+1
= e
hA
e
n

n+1
+h
s

i=1
b
i
(hA) J
n
e
n,i
+h
s

i=1
b
i
(hA)

d
n,i
.
By induction, we derive that for all n {0, . . . , N 1},
e
n+1
=
n

p=0
e
phA

n+1p
(3.20)
+h
s

i=1
b
i
(hA)
n

p=0
e
phA
J
np
(e
np,i
+
np,i
) (3.21)
h
s

i=1
b
i
(hA)
n

p=0
e
phA
J
np

np,i
(3.22)
+h
s

i=1
b
i
(hA)
n

p=0
e
phA

d
np,i
. (3.23)
Since the underlying RungeKutta method is of order s +2, one deduces, as in the proof of Theorem 2.8 that there exists
a positive constant C
1
such that for all h and n,
_
_
_
_
_
n

p=0
e
phA

n+1p
_
_
_
_
_
H
r
C
1
h
s+2
_
_
_
f
(s)
(0)
_
_
H
r+4
+
_
_
f
(s+1)
(0)
_
_
H
r+2
+
T
_
0
_
_
f
(s+1)
()
_
_
H
r+4
d +
T
_
0
_
_
f
(s+2)
()
_
_
H
r+2
d
_
.
(3.24)
To bound the error term (3.21), one uses (3.13) and then (3.15) to derive that
e
np,i
+
np,i

H
r e
np

H
r +ChL
s

j=1
e
np, j

H
r (1 +2Ch
0
Ls)e
np

H
r +2CLh
s

j=1

np, j

H
r .
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.15 (1-19)
G. Dujardin / Applied Numerical Mathematics () 15
Using Hypothesis 3.11 and estimate (3.18), we derive that
3
_
_
_
_
_
h
s

i=1
b
i
(hA)
n

p=0
e
phA
J
np,i
(e
np,i
+
np,i
)
_
_
_
_
_
H
r
Chs
_
n

p=0
L
1
_
(1 +2Ch
0
Ls)e
np

H
r +2CLh
s

j=1

np, j

H
r
__
ChsL
1
(1 +2Ch
0
Ls)
n

p=0
e
p

H
r +2C
2
LL
1

s
h
s+2
T
_
0
_
_
f
(s)
()
_
_
H
r
d.
Hence, there exists positive constants C
2
and C
3
such that for all h and n, the H
r
-norm of the term (3.21) is bounded by
C
2
h

n
p=0
e
p

H
r + C
3
h
s+2
.
In order to bound the H
r
-norm of the error term (3.22), we set

i,s+1
(z) =
s+1
(c
i
z)c
s+1
i

1
s!
s

j=1
a
i, j
(z)c
s
j
,
and write

np,i
=h
s+1

i,s+1
(hA) f
(s)
(t
np
) +

np,i
,
to get
h
s

i=1
b
i
(hA)
n

p=0
e
phA
J
np

np,i
= h
s

i=1
b
i
(0)
n

p=0
e
phA
J
np
h
s+1

i,s+1
(0) f
(s)
(t
np
) (3.25)
h
s

i=1
b
i
(hA)
n

p=0
e
phA
J
np
h
s+1
_

i,s+1
(hA)
i,s+1
(0)
_
f
(s)
(t
np
) (3.26)
h
s

i=1
_
b
i
(hA) b
i
(0)
_
n

p=0
e
phA
J
np
h
s+1

i,s+1
(0) f
(s)
(t
np
) (3.27)
+h
s

i=1
b
i
(hA)
n

p=0
e
phA
J
np

np,i
. (3.28)
Since the underlying RungeKutta method is of order at least s +2, we derive that
s

i=1
b
i
(0)
i,s+1
(0) = 0,
and therefore the term (3.25) reads
h
s+2
n

p=0
e
phA
J
np
_
s

i=1
b
i
(0)
i,s+1
(0)
_
f
(s)
(t
np
) = 0.
On the other hand, there exists s holomorphic functions (
i,s+1
)
i{1,...,s}
such that
i {1, . . . , s}, z C,
i,s+1
(z)
i,s+1
(0) = z
i,s+1
(z).
Note that the functions (
i,s+1
)
i{1,...,s}
are bounded on R. Let M > 0 be a bound for these functions. For all i {i, . . . , s},
an Abel summation yields
3
See the denition of
s
in Theorem 3.6.
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.16 (1-19)
16 G. Dujardin / Applied Numerical Mathematics ()
n

p=0
e
phA
J
np

i,s+1
(hA)(hA) f
(s)
(t
np
)
=
_
n

p=0
e
phA
_
J
0

i,s+1
(hA)(hA) f
(s)
(0) +
n1

p=0
_
p

q=0
e
qhA
_ t
np
_
t
np1
d
dt
_
g
u
_
t, u(t)
_

i,s+1
(hA)(hA) f
(s)
(t)
_
dt.
On one hand,
_
_
_
_
_
_
n

p=0
e
phA
_
J
0

i,s+1
(hA)(hA) f
(s)
(0)
_
_
_
_
_
H
r
T L
1
M
_
_
f
(s)
(0)
_
_
H
r+2
.
On the other hand, since after differentiating and using Hypotheses 3.11 and 3.13, one has
_
_
_
_
d
dt
_
g
u
_
t, u(t)
_

i,s+1
(hA)(hA) f
(s)
(t)
__
_
_
_
H
r
hL
3
M
_
_
f
(s)
(t)
_
_
H
r+2
+hL
3
M

R
_
_
f
(s)
(t)
_
_
H
r+2
+hL
1
M
_
_
f
(s+1)
(t)
_
_
H
r+2
,
where

R denotes the maximum of u

(t)
H
r on [0, T ] (see Hypothesis 3.3), we deduce that there exists a positive constant
C
4
such that for all h and n, the term (3.26) has an H
r
-norm bounded by C
4
h
s+2
.
The same kind of calculations yields, denoting (
i
)
i{1,...,s}
the holomorphic functions such that
i {1, . . . , n}, z C, b
i
(z) b
i
(0) = z
i
(z) =
i
(z)z,
s

i=1
_
b
i
(hA) b
i
(0)
_
n

p=0
e
phA
J
np

i,s+1
(0) f
(s)
(t
np
)
=
i,s+1
(0)
_
n

p=0
e
phA
_

i
(hA)(hA) J
0
f
(s)
(0)
+
i,s+1
(0)
n1

p=0
_
p

q=0
e
qhA
_

i
(hA)(hA)
t
np
_
t
np1
d
dt
_
g
u
_
t, u(t)
_
f
(s)
(t)
_
dt.
If we also denote by M a positive constant bounding the functions (
i
)
i{1,...,s}
, then
_
_
_
_
_
e
qhA

i
(hA)(hA)
t
np
_
t
np1
d
dt
_
g
u
_
t, u(t)
_
f
(s)
(t)
_
dt
_
_
_
_
_
H
r
Mh
t
np
_
t
np1
_
_
_
_
A
d
dt
_
g
u
_
t, u(t)
_
f
(s)
(t)
__
_
_
_
H
r
dt
Mh
t
np
_
t
np1
_
_
_
_
d
dt
_
g
u
_
t, u(t)
_
f
(s)
(t)
__
_
_
_
H
r+2
dt.
Hypotheses 3.11 and 3.13 ensure that
_
_
_
_
d
dt
_
g
u
_
t, u(t)
_
f
(s)
(t)
__
_
_
_
H
r+2

L
3
(1 +

R)
t
np
_
t
np1
_
_
f
(s)
(t)
_
_
H
r+2
dt +

L
1
t
np
_
t
np1
_
_
f
(s+1)
(t)
_
_
H
r+2
dt,
if

R stands for a bound of u

(t)
H
r+2 on [0, T ] (see Hypothesis 3.14). Hence, the H
r
-norm of term (3.27) is bounded by
C
5
h
s+2
for some positive constant C
5
.
One can easily check by using Hypothesis 3.11 that there exists a positive constant C
6
such that for all h and n, the
H
r
-norm of term (3.28) is bounded by C
6
h
s+2
. Hence, term (3.22) is bounded for all h and n by (C
4
+ C
5
+ C
6
)h
s+2
.
Eventually, one can estimate term (3.23) by writing, for i {1, . . . , s}, h (0, h
0
), n {0, . . . , N 1}, and p {0, . . . , n},
using Hypothesis 3.12

d
np,i

H
r L
2
__
_
u(t
np
) u
np,i
_
_
H
r
+
_
_
u(t
np
) u(t
np
+c
i
h)
_
_
H
r
_
e
np,i

H
r .
Therefore, we get that
_
_
_
_
_
h
s

i=1
b
i
(hA)
n

p=0
e
phA

d
np,i
_
_
_
_
_
H
r
CL
2
h
n

p=0
s

i=1
_
e
np,i

2
H
r +

Rhe
np,i

H
r
_
.
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.17 (1-19)
G. Dujardin / Applied Numerical Mathematics () 17
Note that we have using (3.15) and (3.18)
s

i=1
e
np,i

2
H
r
_
s

i=1
e
np,i

H
r
_
2
(s +1)
_
4s
2
e
np

2
H
r +4
s

i=1

np,i

2
H
r
_
(s +1)
_
4s
2
e
np

2
H
r +4s
2
s
h
2s+1
t
np+1
_
t
np
_
_
f
(s)
()
_
_
2
H
r
d
_
(s +1)
_
4s
2
Re
np

H
r +4s
2
s
h
2s+1
t
np+1
_
t
np
_
_
f
(s)
()
_
_
2
H
r
d
_
.
We derive that there exists a positive constant C
7
such that for all h and n, the H
r
-norm of term (3.23) is bounded by
C
7
h
n

p=0
e
p

H
r + C
7
h
s+2
.
The conclusion follows using Lemma 3.7. 2
4. Numerical experiments
We provide two kinds of numerical experiments to illustrate our results. Both of them have been performed with s = 2
collocation points.
The rst method we choose is dened by c
1
=
1
2
and c
2
= 1. This method does not satisfy relation (2.11). Hence, its
underlying RungeKutta method exactly is of order 2.
The other method we choose is dened by c
1
=
1
3
and c
2
= 1. This method satises relation (2.11) but not relation (2.21).
Hence, its underlying RungeKutta method exactly is of order 3.
4.1. Linear problems
Firstly, we consider a linear problem (2.4) with dimension d = 1. The functions u
0
and f are chosen in such a way that
the exact solution of the problem is u(t, x) = e
(t/2)
2
sin(x)
. We set the nal time T = 2. We apply Methods 1 and 2 to this
problem for different time steps h > 0 such that T /h is an integer. We plot in logarithmic scales the L
2
-norm of the nal
error e
T /h
as a function of h on Fig. 1. Calculations are carried out with Fast Fourier Transforms with 2
8
modes.
One can see that, in both cases, the error plot lies between two different straight lines with same slope. The upper line
is reached when T /h = 2/h is close to the square of an integer, that is to say when the time step is resonant (for example,
when T /h = 2/10
1.2019
100 = 10
2
, relation (2.17) does not hold). In this case, inequality (2.12) of Theorem 2.4 holds
and seems to be sharp. On the other hand, for non-resonant time steps, estimate (2.19) of Theorem 2.6 holds and seems to
be sharp. Of course, estimate (2.12) also holds, but it does not seem to be sharp anymore.
4.2. Semi-linear problems
Secondly, we consider a semi-linear problem (3.1) with dimension d = 1. The non-linearity is g(u) = i|u|
2
u and the
initial datum is u
0
(x) = (1 +e
2ix
)/(2 +sin(x)). We set the nal time T = 2. We apply Methods 1 and 2 to this problem
for different time steps h > 0 such that T /h is an integer. We plot in logarithmic scales the L
2
-norm of the nal error e
T /h
as a function of h on Fig. 2. Calculations are carried out with Fast Fourier Transforms with 2
8
modes.
One can see that no resonance occurs for time steps between 10
3
and 10
1
, even when T /h is the square of an
integer, and that both methods have a numerical order consistent with Theorems 3.6 and 3.8: the upper straight line has a
numerical slope close to 2, while the lower straight line has a numerical slope close to 3.
5. Conclusion
This paper provides a numerical analysis of exponential RungeKutta methods of collocation type applied to the linear
and semi-linear Schrdinger equations on a d-dimensional torus. These methods have been studied in [14] when applied to
parabolic problems.
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.18 (1-19)
18 G. Dujardin / Applied Numerical Mathematics ()
Fig. 1. Final error as a function of the time step for a linear Schrdinger problem. Method 1 (upper dotted line) and Method 2 (lower starred line).
Logarithmic scales.
Fig. 2. Final error as a function of the time step for the cubic Schrdinger equation. Method 1 (upper dotted line, numerical slope 1.9932) and Method 2
(lower starred line, numerical slope 2.9874). Logarithmic scales.
This paper shows how the results of [14] extend to the case of the Schrdinger equation and provides sucient con-
ditions to achieve orders s, s + 1 and s + 2 using an s-stage method (see Theorems 2.1, 2.4, 2.6, 2.8 and 2.9 for linear
Schrdinger problems and Theorems 3.6, 3.8 and 3.15 for non-linear Schrdinger problems). Essentially, when solving
Schrdinger problems, the methods behave as they do when solving parabolic problems: under some suitable assump-
tions (on the right-hand side or on the non-linearity), an s-stage method is of order s and can achieve orders s + 1 and
s + 2. However, the proofs of the results presented in this paper require some hypotheses (for example on the regularity
of the exact solution of the problem, like Hypotheses 3.3 and 3.14) that seem more restrictive in the Schrdinger context
than the corresponding ones in the parabolic context. Moreover, this paper points out a major difference between parabolic
problems and Schrdinger problems solved by such exponential RungeKutta methods: it shows and explains the effect of
resonant and non-resonant time steps over nite time intervals when solving linear Schrdinger problems (see Theorems
2.6 and 2.9).
ARTICLE IN PRESS APNUM:2245
Please cite this article in press as: G. Dujardin, Exponential RungeKutta methods for the Schrdinger equation, Applied Numerical Mathematics (2009),
doi:10.1016/j.apnum.2009.02.002
JID:APNUM AID:2245 /FLA [m3G; v 1.24; Prn:10/03/2009; 15:16] P.19 (1-19)
G. Dujardin / Applied Numerical Mathematics () 19
Acknowledgements
The author would like to thank Philippe Chartier and Erwan Faou for their comments and ideas about this work.
References
[1] A. Alinhac, P. Grard, Oprateurs pseudo-diffrentiels et thorie de NashMoser, Interditions/ditions du CNRS, 1991.
[2] H. Berland, A.L. Islas, C.M. Schober, Conservation of phase space properties using exponential integrators on the cubic Schrdinger equation, J. Comput.
Phys. 255 (1) (2007) 284299.
[3] H. Berland, B. Skaestad, Solving the nonlinear Schrdinger equation using exponential integrators, in: Proceedings SIMS 2005 46th Conference on
Simulation and Modeling, Trondheim, Norway.
[4] D. Cohen, E. Hairer, C. Lubich, Long-time analysis of nonlinearly perturbed wave equations via modulated Fourier expansions, Arch. Ration. Mech.
Anal. 187 (2008) 341368.
[5] D. Cohen, E. Hairer, C. Lubich, Conservation of energy, momentum and actions in numerical discretizations of nonlinear wave equations, April 2007,
preprint.
[6] G. Dujardin, tude de schmas de discrtisation en temps de lquation de Schrdinger, PhD thesis, 2008.
[7] G. Dujardin, E. Faou, Normal form and long time analysis of splitting schemes for the linear Schrdinger equation with small potential, Numer.
Math. 106 (2) (2007) 223262.
[8] B. Garca-Archilla, J. Sanz-Serna, R. Skeel, Long-time-step methods for oscillatory differential equations, SIAM J. Sci. Comput. 30 (3) (1998) 930963.
[9] V. Grimm, M. Hochbruck, Error analysis of exponential integrators for oscillatory second-order differential equations, J. Phys. A, Math. Gen. 39 (2006)
54955507.
[10] E. Hairer, C. Lubich, Spectral semi-discretisations of weakly nonlinear wave equations over long times, Found. Comput. Math. 8 (2008) 319334.
[11] E. Hairer, C. Lubich, G. Wanner, Geometric Numerical Integration. Structure-preserving Algorithms for Ordinary Differential Equations, second ed.,
Springer, Berlin, 2006.
[12] M. Hochbruck, C. Lubich, A Gautschi-type method for oscillatory second-order differential equations, Numer. Math. 83 (1999) 403426.
[13] M. Hochbruck, A. Ostermann, Explicit exponential RungeKutta methods for semilinear parabolic problems, SIAM J. Numer. Anal. 43 (3) (2005) 1069
1090.
[14] M. Hochbruck, A. Ostermann, Exponential RungeKutta methods for parabolic problems, Appl. Numer. Math. 53 (24) (2005) 323339.
[15] B. Minchev, W.M. Wright, A review of exponential integrators for semilinear problems, Tech. rep. 2/05, Department of Mathematical Sciences, Norwe-
gian University of Science and Technology, Trondheim, Norway.
[16] J.M. Sanz-Serna, Modied impulse methods for highly oscillatory differential equations, SIAM J. Numer. Anal. 46 (2) (2008) 10401059.