Multiscale Splitting Method for the

BoltzmannPoisson Equation: Application to the

Dynamics of Electrons

J
urgen Geiser

Ernst-Moritz-Arntz University of Greifswald,

Department of Physics,
Domstr. 14, D-17487 Greifswald, Germany
juergen.geiser@uni-greifswald.de

Abstract. In this paper we present a model based on dynamics of elec-

trons in a plasma using a simplified Boltzmann equation coupled with
Poissons equation. The motivation arose from simulating active plasma
resonance spectroscopy, which is used for plasma diagnostic techniques,
see [1], [15] and [16]. Mathematically, we are interested in designing split-
ting methods for the model problem. While the full Boltzmann equation
is delicate to solve, we decouple it into a transport and collision part,
which are then solved in different ways. First we reduce to a simplified
transportcollision equation and start to analyse the abstract Cauchy
problem using semigroup methods. Second, we pass to the coupled trans-
port and collision model and apply the splitting ideas, resecting the differ-
ent discretisation schemes. The results are discussed first with numerical
experiments and then we verify the underlying theoretical novelties.

Keywords: kinetic model, neutron transport, dynamics of electrons, trans-

port equation, splitting schemes, semigroup.

AMS subject classifications. 35K25, 35K20, 74S10, 70G65.

1 Introduction
Our motivate arose from studying the simulation of active plasma resonance
spectroscopy, a well established plasma diagnostic technique. To study this tech-
nique with simulation models, we concentrate on an abstract kinetic model that
describes the dynamics of the electrons in a plasma by using a Boltzmann equa-
tion. The Boltzmann equation is coupled with the electric field and we obtain
coupled partial differential equations. The full Boltzmann equation is a very del-
icate equation to solve. Therefore, we decouple this into transport and collision
contributions, which are solved in different ways. While finite difference schemes
are applied to the transpoort parts, the collision part is solved with numerical
integration schemes. The underlying splitting scheme is theortically discussed as
an abstract Cauchy problem. Here, we could address the Cauchy problem with
2

functional analytical tools. First, we will discuss the description of a positive

semigroup, which helps to carry out the numerical estimates in the splitting
schemes. Second, a numerical method is discussed with respect to separate dif-
ferential and integral parts of the equations. The numerical approximation of
the abstract splitting scheme is made by applying an iterative splitting method
of the second order.
The paper is outlined as follows. In Section 2, we present our mathematical model
and a possible reduced model for further approximations. The functional analyt-
ical setting with semigroups is discussed in Section 3. The splitting schemes are
presented in Section 4. The results of some numerical experiments are exhibited
in Section 6. In the conclusions that are given in Section 7, we summarize our
results.

2 Mathematical Model

In the following a model is presented whose physical motivation is explained in

[1], [15] and [16].
The kinetic model considers a fluid dynamical approach to treat the natural
ability of plasmas to resonate near the electron plasma frequency pe .
Here we specialise to an abstract kinetic model to describe the dynamics of
the electrons in the plasma which allows this resonance.
The Boltzmann equation for the electrons is:

f (x, v, t)
= A[f (x, v, t)] + B[f (x, v, t)], (x, v) x v , t [0, T ], (1)
t
e
A[f ] = v x f (x, v, t) x v f (x, v, t), (2)
me
Z
B[f ] = (x, v, t)f (x, v, t) + (x, v, v )f (x, v , t) dv , (3)
V
f (x, v, 0) = f0 (x, v), (x, v) P,x P,v , (4)
f (x, v, t) = f1 (x, v), (x, v) P,x P,v , (5)

where P,x P,v IR3 IR3 is the six-dimensional phase space. f : P,x
P,v [0, T ] Xc is the density function and Xc is an appropriate smooth
space, e.g. C2 . f0 is the initial function and the f1 is the boundary condition in
the plasma phase space P,x P,v .
We assume, based on the different materials (a plasma and a dielectricum) a
complete reflection of the electrons due to the sheath f (v|| + v ) with v|| , which
is the parallel and v perpendicular to the surface normal vector. Further the
electric force-field is given as F = ex , where : P,x Xc is the potential
and Xc is an appropriate smooth space, e.g. C2 .
Boltzmanns equation is coupled with the electric field , while the potential
is approximated via Poissons equation. The electrostatic approximation of the
field is represented by a potential that is valid on the complete velocity volume
 3

S P,v and apply Poissons equation:

 R
e(ni (x) f dS) in x P,x
x () = , (6)
0 in x D,x
where the full space domain is x = P,x D,x IR3 and the full velocity
domain is v = P,v D,v . The permittivity is equal to 0 in the plasma P,x
and 0 D in the dielectric D,x . fulfills the boundary conditions Un at any
electrode En and n = 0 at isolator I, whereas n is the normal vector of
the isolator surface.
On the surface of the dielectric, a surface charge may accumulate, which
leads to a transition condition:
() = . (7)
We will make the following assumption to modify the model into a simpler and
more tractable system of equations, which allows linearizing and simplifying the
Boltzmann equations, see [20].
First we assume we have an analytical solution of the potential , such that
we could reset to a diffusive term as:
F
v x Dx , (8)
me
where D is the diffusion tensor and F = e is the external force. This modified
model allows employing the following semi-group theory and deriving splitting
schemes for the model.
The modified and splittable model equation is
f (x, v, t)
= A[f (x, v, t)] + B[f (x, v, t)], (x, v) x v , t [0, T ], (9)
t
A[f ] = v x f (x, v, t) x Dx f (x, v, t), (10)
Z
B[f ] = (x, v, t)f (x, v, t) + (x, v, v )f (x, v , t) dv , (11)
V
f (x, v, 0) = f0 (x, v), (x, v) P,x P,v , (12)
f (x, v, t) = f1 (x, v), (x, v) P,x P,v . (13)
Next, we will discuss the relevant semigroup theory.

3 Semigroups for Transport Equations

In the following, we derive the exponential growth of the transport semigroups
that are used in the section on the numerical methods.
We discuss two aspects:
Neutron transport,
Electron transport.
The difference of the two transport schemes are:
Neutron transport is not influenced by an electric field,
Electron transport is influenced by an electric field.
4

3.1 Transport model for the neutrons

For this model we can assume that f (x, v, t) describes the density distribution
of the particles at position x S with speed v V at time t [0, T ], see also
[3] and [17].
The space S is assumed to be a compact and convex subset of IR3 with
nonempty interior, and the velocity space V is
V := {v IR3 : vmin ||||2 vmax }
for vmin > 0 and vmax < .

Assumption 31 We make the following assumptions:

Particles move according to their speed v.

Particles are absorbed according to the probability given by the function
which depends on x and v.
Particles are scattered according to a scattering kernel depending on x, the
incoming speed v , and the outgoing speed v.

Then the neutron transport is

f (x, v, t)
= v f (x, v, t) (x, v, t)f (x, v, t) (14)
Z t
+ (x, v, v )f (x, v , t) dv ,
V
f (x, v, 0) = f0 (x, v), (15)

and boundary conditions are included in the transport operator A0 see, in the
following, the abstract Cauchy problem.
We next treat the abstract Cauchy problem for this simplified model.

Abstract Cauchy problem: Transport model for the neutrons We have

a Banach space X := L1 (S V ) with Lebesgue measure on S V IR6 and
define the abstract Cauchy problem as:
du(t)
= Bu(t) , (16)
dt
du(t)
= (A0 M + K )u(t) , (17)
dt
u(0) = u0 , (18)

where u X.
We have the following operators:
1.) Collisionless transport operator
2.) Absorption operator
3.) Scattering Operator
Important results for the further numerical analysis is the fact that the trans-
port semigroup can be estimated by an exponential growth, see [3]:
 5

Corollary 1. We assume that s(B) > is a dominant eigenvalue and (S(t))t0

is irreducible. Then the transport semigroup (S(t))t0 has balanced exponential
growth. There exists a one-dimensional projection P satisfying 0 < P f whenever
0 < f such that:

|| exp(s(B)t)S(t) P || M exp(t), (19)

for all t 0 and appropriate M 1 and > 0.

3.2 Transport model for the electrons or ions

For this model we can assume that f (x, v, t) describes the density distribution
of particles at position x S with speed v V at time t [0, T ], see also [3]
and [17].
The space S is assumed to be a compact and convex subset of IR3 with
nonempty interior, and the velocity space V is
V := {v IR3 : vmin ||||2 vmax }
for vmin > 0 and vmax < .

Assumption 32 We make the following assumptions:

Particles move according to their speed v.

Particles are absorbed according to the probability given by the function
which depends on x and v.
Particles are scattered according to a scattering kernel depending on x, the
incoming speed v , and the outgoing speed v.
Particles are influenced by the static electric field , which can be derived by
the kinetic theory.

The electron transport is

f (x, v, t) e
= v x f (x, v, t) x v f (x, v, t)
t me
Z
(x, v, t)f (x, v, t) + (x, v, v )f (x, v , t) dv , (20)
V
f (x, v, 0) = f0 (x, v), (21)

and boundary conditions are included in the transport operators. is the electric
field.
Furthermore, we have Poissons equation,

 R
e(ni f dS) in P
x () = , (22)
0 inD

the permittivity is equal to 0 in the plasma P and 0 D in the dielectric D.

6

For the simplification, we assume to solve the Poissons equation analytically

with:
1 P e(ni f dS)dx dx in P
Z R R R
= D(f )dx = P , (23)
0 inD
where D is the analytical function of the electric field x .
We embed the electric field analytically into the transport equation.
f (x, v, t)
= v x f (x, v, t) x D(f ) x f (x, v, t)
t Z
(x, v, t)f (x, v, t) + (x, v, v )f (x, v , t) dv , (24)
V
f (x, v, 0) = f0 (x, v), (25)
and boundary conditions are included in the transport operators A0 and A1 ,
see, in the following, the treatment of the abstract Cauchy problem. D(f ) is the
diffusion parameter that includes the electric field and we assume to approximate
via a constant operator D D(f ).
Next we treat the abstract Cauchy problem for a transport model for the
electrons or ions.

Abstract Cauchy problem: Transport model for the electrons or ions.

We have a Banach space X := L1 (S V ) with Lebesgue measure on S V IR6
and define the abstract Cauchy problem as
du(t)
= Bu(t) , (26)
dt
du(t)
= (A0 + A1 M + K )u(t) , (27)
dt
u(0) = u0 , (28)
where u X.
We have the operators
1.) Collisionless transport operator
2.) Diffusion operator
3.) Absorption operator
4.) Scattering Operator
An important result for the further numerical analysis is the fact that the
transport semigroup can be estimated by an exponential growth. Due to the
analytical embedding of the electric field, we could also estimate such operator.
Corollary 2. We assume that s(B) > is a dominant eigenvalue and (S(t))t0
is irreducible Then the transport semigroup (S(t))t0 has balanced exponential
growth. There exists a one-dimensional projection P satisfying 0 < P f whenever
0 < f such that:

|| exp(s(B)t)S(t) P || M
exp(t), (29)
1 and > 0.
for all t 0 and appropriate M
 7

In the next section we discuss the splitting schemes.

4 Splitting Schemes

In general, operator splitting methods are used to solve complex models in geo-
physical and environmental physics. They have been developed and applied in
[19], [21] and [22].

4.1 Sequential splitting method for nonlinear problems

For our problems, nonlinear splitting schemes are necessary, see [13]. We could
use the result for the general formulation of nonlinear ordinary differential equa-
tions, which are given by

c (t) = F1 (t, c(t)) + F2 (t, c(t)) , (30)

where the initial conditions are cn = c(tn ).

As before, we can decouple the above problem into two (usually simpler)
sub-problems, namely,

c (t)
= F1 (t, c (t)) with tn t tn+1 and c (tn ) = cn , (31)
t
c (t)
= F2 (t, c (t)) with tn t tn+1 and c (tn ) = c (tn+1 ) , (32)
t
where the initial values are given by cn = c(tn ) and the split approximation on
the next time level is defined as cn+1 = c (tn+1 ).
For this case, the splitting error can be determined by using the Jacobians
of the non-linear mappings F1 and F2 :

1 F1 F2
n = [ F2 , F1 ](tn , c(tn )) + O(n2 ) . (33)
2 c c
Hence, for the general case, the splitting error is of first order, i.e., O(n ).

Remark 1. Higher order splitting methods are given in [10]. Based on the Strang
splitting, higher order nonlinear splitting methods are also possible, see [2].

In the next subsection we present the iterative splitting method.

4.2 Iterative splitting method

Alternatives are iterative approaches to nonlinear splitting schemes.

We concentrate again on nonlinear differential equations of the form

du
= A(u(t))u(t) + B(u(t))u(t), with u(tn ) = un , (34)
dt
8

where A(u), B(u) are matrices with nonlinear entries and densely defined, where
we assume that the entries involve the spatial derivatives of c, see [23]. In the
following we discuss the standard iterative operator splitting method as a fixed-
point iteration method to linearize the operators.
We split our nonlinear differential equation (34) by applying
dui (t)
dt = A(ui1 (t))ui (t) + B(ui1 (t))ui1 (t), with ui (tn ) = cn , (35)
dui+1 (t) n n
dt = A(ui1 (t))ui (t) + B(ui1 (t))ui+1 (t), with ui+1 (t ) = c , (36)

where the time step is = tn+1 tn . The iterations are i = 1, 3, . . . , 2m + 1.

u0 (t) = cn is the starting solution, where we assume that the solution cn+1 is
near cn , or u0 (t) = 0. Thus we have to solve the local fixed-point problem. cn is
the known split approximation at time level t = tn .
The split approximation at time level t = tn+1 is defined as cn+1 = u2m+2 (tn+1 ).
We assume that the operators A(ui1 (tn+1 )), B(ui1 (tn+1 )) are constant for
i = 1, 3, . . . , 2m + 1. Here the linearization is done with respect to the iterations,
such that A(ui1 ), B(ui1 ) are at least non-dependent operators in the iterative
equations, and we can apply the linear theory. For the linearization we assume
at least in the first equation A(ui1 (t)) A(ui (t)), and in the second equation
B(ui1 (t)) B(ui+1 (t)), for small t.
We have
||A(ui1 (tn+1 ))ui (tn+1 ) A(un+1 )u(tn+1 )|| ,
for sufficient iterations i {1, 3, . . . , 2m + 1}.

Remark 2. The linearization with the fixed-point scheme can be used for smooth
or weakly nonlinear operators, but otherwise we lose the convergence behavior,
not converging to a local fixed point, see [14].

5 Numerical Integration of the Integro-Part

We treat the following integro-differential equation:

Z t
u
= u(s) ds, (37)
t 0
u(0) = u0 , (38)

The integration part is done numerically with

Trapezoidal rule:

b n1  !
ba f (a) + f (b) X ba
Z
f (x) dx + f a+k (39)
a n 2 n
k=1

where the subintervals have the form [kh, (k + 1)h], with h = (ba)/n and k =
0, 1, 2, ..., n1.
 9

Degree Common name Formula Error term

ba (ba)3 (2)
1 Trapezoid rule 2
(f0 + f1 ) 12
f ()
ba (ba)5 (4)
2 Simpsons rule 6
(f 0 + 4f1 + f2 ) 2880
f ()
ba (ba)5 (4)
3 Simpsons 3/8 rule 8
(f0 + 3f1 + 3f2 + f3 ) 6480 f ()
(ba)7
4 Booles rule ba
90
(7f0 + 32f 1 + 12f2 + 32f 3 + 7f4 ) 1935360
f (6) ()
Table 1. Numerical integration formula (closed NewtonCotes formula)

The higher order formulas are given as closed NewtonCotes formulas:

where fi is a shorthand for f (xi ), with xi = a + i(b a)/n, and n the degree.
We obtain the following formula for the Trapezoid rule,

u
= t/2(u(0) + u(t)) ds, (40)
t
u(0) = u0 , (41)

and obtain the analytical result

2 t2 1
u(t) = exp( )u(0) u(0), (42)
2 4 2

For a higher order formula like Simpsons rule, we have

u
= t/6(u(0) + 4u(t/2) + u(t)) ds, (43)
t
u(0) = u0 . (44)

We apply the idea of a polynomial solution:

u(t) = a0 + a1 t + a2 t2 + a3 t3 + . . .
and we obtain, after differentiating the coefficients,

a1 + 2a2 t + 3a3 t2 + . . . (45)

= t/6 a0 + 4(a0 + a1 t/2 + a2 t2 /4 + a3 t3 /8 + . . .)
+a0 + a1 t + a2 t2 + a3 t3 + . . . ,

a0 = u 0 . (46)

Now, comparing coefficients yields

a0 = u 0 (47)
a1 = a3 = a5 = . . . = 0 (48)
a2 = 3a0 (49)
1
a4 = a2 , . . . . (50)
12
10

6 Numerical Experiments
We present the results of our numerical experiments based on the neutron trans-
port. A simplified one-dimensional model is given by
Z
t c + vx c Dxx c + c = (x, v, v )c(x, v , t) dv .

The velocity v and the diffusion D are given by the plasma model. The initial
conditions are given by c(x, 0) = c0 (x) and the boundary conditions are trivial,
n c(x, t) = 0.
A first integral operator is
Z Z T

(x, v, v )c(x, v , t) dv = c(x, t)dt.
0

A second integral operator is

We assume a simple collision operator: (x, v, v ) = q(v )(1 + v 2 ),
where q(v ) is the potential, e.g., v 2 .

We deal with the first integral operator and define the following operators:
1 1
A = v x [1 1 0]I D x 2 [1 2 1]I

B = ( + t)I,
while
exp(Bt) = exp((t + t2 /2)I),
where I is the identity matrix.
In the following, the simplified real-life problem for a neutron transport equa-
tion, which includes the gain and loss of a neutron, will be presented.
We concentrate on the computational benefits of a fast computation of the
iterative scheme, given with matrix exponentials.

The equation is
Z t
t c + Fc = 1 c + 2 c(x, t)dt, in [0, t], (51)
0
F = v D, (52)
c(x, t) = c0 (x), on , (53)
c(x, t) = c1 (x, t), on [0, t]. (54)

In the following, we deal with the semidiscretized equation given by the

matrices

t C = (A 1 + 2 ) C, (55)

where C = (c1 , . . . , cI )T is the solution of the species in the mobile phase in each
spatial discretization point (i = 1, . . . , I).
 11

We have the following two operators for the splitting method:

2 1
1 2 1
D .. .. ..

A= (56)

x2
. . .

1 2 1
1 2

1
1 1
v .. ..

+ IRII (57)

x . .
1 1
1 1
where I is the number of spatial points.

1 0
0 1 0

1 =
. . .
.. .. .. IRII

(58)

0 1 0
0 1
For the integral term we have the following ideas:
Case
Rt 1:
0
2 c(x, t)dt 2 tc(x, t),
and we obtain the matrix
2 t2 /2

0

0 2 t2 /2 0

2 =
. .
.. .. . .. IRII .

(59)

0 2 t2 /2 0
0 2 t2 /2
For the operator splitting scheme, we apply A and B = 1 + 2 and we
apply the iterative splitting method, given in Equations (35)(36).
Case 2:
We integrate the operator B with respect to the previous solutions Ci1
2 (Ci1 ) and we obtain the matrix
2 (Ci1 )
R t
2 c 1,i1 (x, s) ds 0 0
0 Rt
0 0 2 c2,i1 (x, s) ds 0


= .. .. .. ..
. . . .


Rt
0 0 0 2 cI,i1 (x, s) ds
IRII . (60)
12

We obtain B(C) = 2 (Ci1 ) + 1 C.

The iterative scheme is given by
For i = 1, 2, . . .

Ci (t) = exp(A(t tn ))C(tn )

Rt (61)
+ tn exp((t s)A)B(Ci1 (s)) ds, t (tn , tn+1 ].

For the reference solution, we apply a fine time- and spatial scale without
decoupling the equations. Figure 1 presents the numerical errors between the
exact and the numerical solution. Here we obtain the optimal results for one-
sided iterative schemes on the operator B, meaning that we iterate with respect
to B and use A as the right hand side.

Remark 3. For all iterative schemes, we can reach faster results than with the
standard schemes, due to the fact that the iterative schemes benefit from their
fast computations of the exponential matrices. With from four to five iterative
steps, we obtain more accurate results than we did with the expensive stan-
dard schemes. With one-sided iterative schemes, we obtain the best convergence
results.

7 Conclusions and Discussion

We presented the coupled model for the transport of deposition species in a

plasma environment. We assumed the flow field could be computed by the plasma
model and the transport of the deposition species by a transportreaction model.
Such a first model can help understand the important modeling of the plasma
environment in a CVD reactor.

References
1. N.St.J. Braithwaite and R.N. Franklin. Reflections on electrical probes. Plasma
Resource Sci. Technol., 18, 014008, 2009.
2. S. Descombes and M. Thalhammer. The LieTrotter splitting method for nonlinear
evolutionary problems involving critical parameters. A compact local error repre-
sentation and application to nonlinear Schrdinger equations in the semi-classical
regime. IMA J. Numer. Anal., doi: 10.1093/imanum/drs021, 2012.
3. K.-J. Engel and R. Nagel, One-Parameter Semigroups for Linear Evolution Equa-
tions. Springer-Verlag, New York, 2000.
4. I. Farago. Splitting methods for abstract Cauchy problems. In: Lect. Notes Comp.
Sci. vol. 3401, Springer-Verlag, Berlin, 2005, pp. 3545.
5. I. Farago, J. Geiser. Iterative Operator-Splitting methods for Linear Problems.
Preprint No. 1043 of the Weierstrass Institute for Applied Analysis and Stochastics,
Berlin, June 2005.
6. J. Geiser. Numerical Simulation of a Model for Transport and Reaction of Ra-
dionuclides. Proceedings of the Large Scale Scientific Computations of Engineering
and Environmental Problems, Sozopol, Bulgaria, 2001.
 13

7. J. Geiser. Decomposition Methods for Partial Differential Equations: Theory and

Applications in Multiphysics Problems. Chapman & Hall/CRC, 2009.
8. J. Geiser. Computing Exponential for Iterative Splitting Methods. Journal of Ap-
plied Mathematics, special issue: Mathematical and Numerical Modeling of Flow
and Transport, accepted, January 2011.
9. I. Glowinski. The iterative operator-splitting methods. Preprint, University of
Houston, 2004.
10. E. Hansen and A. Ostermann. Exponential splitting for unbounded operators.
Mathematics of Computation, 78:14851496, 2009.
11. W.H. Hundsdorfer. Numerical Solution of AdvectionDiffusionReaction Equa-
tions. Technical Report NM-N9603, CWI, 1996.
12. W.H. Hundsdorfer, J. Verwer W. Numerical Solution of Time-dependent
AdvectionDiffusionReaction Equations, Springer-Verlag, Berlin, 2003.
13. J. Geiser. Iterative Splitting Methods for Differential Equations. Chapman &
Hall/CRC, 2011.
14. J. Kanney, C. Miller and C. Kelley. Convergence of iterative split-operator ap-
proaches for approximating nonlinear reactive transport problems. Advances in
Water Resources, 26:247261, 2003.
15. M. Lapke, T. Mussenbrock and R.P. Brinkmann. Modelling of volume- and surface
wave based plasma resonance spectroscopy. Abstracts IEEE International Confer-
ence on Plasma Sciences, pp. 89, 2010.
16. J. Oberrath, M. Lapke, T. Mussenbrock and R.P. Brinkmann. A Functional Ana-
lytical Description of Active Plasma Resonance Spectroscopy in Terms of Kinetic
Theory. Proceedings of the 30th ICPIG, Belfast, August 28September 2, 2011.
17. A. Rhandi. Spectral Theory for Positive Semigroups and Applications. Quaderno
Q. 1-2002, University of Lecce, Italy, 2002.
18. T.K. Senega and R.P. Brinkmann. A multi-component transport model for
non-equilibrium low-temperature low-pressure plasmas. J. Phys. D: Appl.Phys.,
39:16061618, 2006.
19. G. Strang. On the construction and comparision of difference schemes. SIAM J.
Numer. Anal., 5:506517, 1968.
20. S. Ukai and T. Yang. Mathematical Theory of Boltzmann Equation. Lecture Notes
Series No 8. Liu Bie Ju Center of Mathematical Sciences, City University of Hong
Kong, 2006.
21. J.,G. Verwer and B. Sportisse. A note on operator splitting in a stiff linear case.
MAS-R9830, 1998.
22. Z. Zlatev. Computer Treatment of Large Air Pollution Models. Kluwer, 1995.
23. E. Zeidler. Nonlinear Functional Analysis and its Applications. II/B Nonlinear
montone operators Springer-Verlag, Berlin, 1990.
14

AB, Strang, oneside with A

1
10

0
10

1
10
c1
2 AB
10
L1

3
10 c2
err

4 Strang
10
c3
5 c6
10

6
10 c4
c5
7
10
2 1 0
10 10 10
t

AB, Strang, oneside with B

2
10

0
10

c1
2
10 AB

c2
errL1

4 Strang
10
c3

6 c4
10

c5
8
10
c6

10
10
2 1 0
10 10 10
t

AB, Strang, twoside

2
10

0
10

c1
2
10 AB
L1

c2
err

4 Strang
10
c3
c6

6
10 c4

c5
8
10
2 1 0
10 10 10
t

Fig. 1. Numerical errors of the one-sided splitting scheme with A (upper figure), the
one-sided splitting scheme with B (middle figure), and the iterative schemes with
1, . . . , 6 iterative steps (lower figure).