High Order Difference Approximations for the

Linearized Euler Equations

Stefan Johansson

Abstract
The computers of today make it possible to do direct simulation of
aeroacoustics, which is very computationally demanding since a very high
resolution is needed.
In the present thesis we study issues of relevance for aeroacoustic sim-
ulations. Paper A considers standard high order difference methods. We
study two different ways of applying boundary conditions in a stable way.
Numerical experiments are done for the 1D linearized Euler equations.
In paper B we develop strictly stable difference methods which give
smaller dispersion errors than standard central difference methods. The
new methods are applied to the 1D wave equation.
Finally in Paper C we apply the new difference methods to aeroacoustic
simulations based on the 2D linearized Euler equations.
Taken together, the methods presented here are strictly stable by con-
struction. They lead to better approximation of the wave number, which
in turn results in a smaller L2 -error than obtained by previous methods
found in the literature. The results are valid when the problem is not
fully resolved, which usually is the case for large scale applications.
Acknowledgments
First, I want to thank my supervisors Bernhard Müller and Mickael Tuné
for all your support, and thoroughness in reviewing my work.
Finally, I wish to thank all the people at TDB for creating a good
atmosphere, and supporting many interesting discussions.

2
List of Appended Papers
This thesis is a summary of the following papers. References to the papers
are made using the capital letter associated with each paper.
Paper A Stefan Johansson. Numerical Solution of the Linearized Euler
Equations Using High Order Finite Difference Operators with the
Summation by Parts Property. Report 2002-034, Department of
Information Technology, Uppsala University, Uppsala, 2002 (revised
version, March 2003).
Paper B Stefan Johansson. High Order Finite Difference Operators with
the Summation by Parts Property Based on DRP Schemes. Report
2004-036, Department of Information Technology, Uppsala Univer-
sity, Uppsala, 2004 Submitted to BIT.
Paper C Stefan Johansson. High Order Summation by Parts Opera-
tor Based on a DRP Scheme Applied to 2D Aeroacoustics. Report
2004-050, Department of Information Technology, Uppsala Univer-
sity, Uppsala, 2004.
Some numerical experiments from Paper A were presented by the au-
thor at the 10th International Congress on Sound and Vibration (2003)
[7], together with work done by Bernhard Müller. Preliminary experi-
ments using the new method from Paper B were included in joint pa-
pers with Bernhard Müller presented by Müller at EUROMECH Collo-
quium 449, Chamonix, France 2003 [6] and in Proceedings of 4th EC-
COMAS Congress 2004, Jyväskylä, Finland [8]. The paper presented at
EUROMECH will appear in Comptes Rendus Mecanique (2004/2005).

3
Contents
Summary
Paper A
Paper B
Paper C

4
 1 Introduction
The development of methods for approximate solution of partial differ-
ential equations accelerated after the second world war due to the rapid
development of computers that made it possible to solve non trivial prob-
lems. Partial differential equations are used by scientists to model wave
propagation, heat transfer, fluid flow around rigid objects and many other
applications.
Numerical analysis for partial differential equations, the subject of
this thesis, is the art of approximating a partial differential equation,
and thereafter solving the resulting algebraic problem. There are several
alternative approaches to do the approximation. We have chosen the
finite difference method, i.e., the variables are represented at grid points
and derivatives are approximated using Taylor expansions. The reason
for choosing finite difference methods is that they are relatively easy to
implement, and in explicit form they are in most cases more efficient for
wave propagation problems than for example the finite element method.
The computers of today make it possible to do direct simulation of
aeroacoustics, which is very computationally demanding since a very high
resolution is needed [13]. The application of aeroacoustics includes simu-
lation of noise emitted from cars and airplanes. An example of such an
application is the study by Johan Westerlund [15] where he studied rocket
launch noise from a simplified model of the Arianne V rocket. A picture
of acoustic pressure just before lift off is shown in Figure 1.

Acoustic pressure Around Ariane V

110

100
100
90

80
80

70

60
x

60

50
40
40

20 30

20

0
−60 −40 −20 0 20 40 60 10
y

Figure 1: Solution for acoustic pressure measured in dB, when the frequency is
33 Hz.

In the present thesis we study issues of relevance for aeroacoustic sim-

ulations. Paper A considers standard high order difference methods. We

5
study two different ways to apply boundary conditions in a stable way.
Numerical experiments are done for the 1D linearized Euler equations.
In paper B we develop strictly stable difference methods which give
smaller dispersion errors than standard central difference methods. The
new methods are applied to the 1D wave equation.
Finally in Paper C we apply the new difference methods to aeroacoustic
simulations based on the 2D linearized Euler equations.

2 Overview of Stability Theory for Finite

Difference Methods
Given a well-posed partial differential equation (PDE) there is no con-
structive way to write down the solution in closed form, and for many
PDEs there are no known solutions in closed form. What can be done is
to construct an approximation of the PDE that can be made arbritrary
good that is ”easy” to solve. One way of doing that is to approximate the
derivatives with finite differences. Examples of finite differences for a grid
function u(t)j = u(t, j ∗ h), where h is the step size are

uj+1 − uj−1 uj − uj−1 uj+1 − uj

D0 uj = D− uj = D+ uj =
2h h h
(1)
that can easily be shown to approach ∂u/∂x when h → 0 by Taylor
expansion.
The three important concepts for finite difference methods are consis-
tency, stability and convergence. Consistency means that the approxima-
tion of the PDE converges to the PDE itself when letting the number of
grid points go infinity, stability that the approximation in some sense is
insensitive to perturbations of the initial data, boundary data and round
of errors and convergence that the approximate solution approaches the
exact solution when the number of grid points goes to infinity.
The connection between consistency, stability and converges is given
by the famous Lax-Richtmyer theorem that says that for a well-posed
linear PDE a given consistent finite difference scheme is stable if and only
if it is convergent.
Consistency of the method almost always follows from its construction,
but the tricky part is to show stability, especially for initial boundary value
problems. We start with the periodic case and let time be continuous,
and look at the simplest hyperbolic PDE ut = ux , with initial condition
u(x, 0) = f . Let us discretize ∂/∂x with D0 . If the problem is periodic
we can use the Fourier transform to investigate stability.

duj uj+1 − uj−1

= (2)
dt 2h
If we use Fourier transform formally on equation (2), the PDE trans-
forms to an ordinary differential equation (ODE). For a wave number ω
the ODE reads

deiωjh ũ 1 ³ iω(j+1)h ´
= e ũ − eiω(j−1)h ũ ⇒
dt 2h
dũ 1 ³ iωh −iωh ´ dũ i
= e e ũ ⇒ = sin(ωh)ũ (3)
dt 2h dt h

6
.
It follows from Parseval’s relation kũk = kuk that the L2 -norm of u is
constant in time, because

i
ũ(ω, t) = e h sin(ωh)t ũ(ω, 0) ⇒ kũ(ω, t)k = kũ(ω, 0)k (4)
⇒ ku(·, t)k
|{z} = ku(·, 0)k (5)
P arseval

In general for the PDE ut = P u the Fourier transform will yield

ũ(ω, t) = eQt ũ(ω, 0), where Q is the Fourier transform of P , and a suffi-
cient condition for stability will be that the eigenvalues of Q have a non
positive real part. For fully discrete approximations the condition is that
the spectral radius of Q is less than or equal to 1 + αk, where α is a
constant independent of x, h, and k and k is the time step. This method
of investigating stability is also called von Neumann analysis.
The energy method can also be used. The idea is to show that the
growth of the norm of u can be estimated independently Pof h.
The discrete scalar product is defined as (u, v)h = N j=0 uj vj h, with
the induced norm kuk2h = (u, u)h . For equation (2) the time derivative of
kuk is

d
kuk = (u, ut )h + (ut , u)h = (u, D0 u)h + (D0 u, u)h =
dt
− 2hu0 u−1 + 2huN uN −1 = 0 (6)

That is the same result as we got using von Neumann analysis.

In general show that the approximation D of P is semi-bounded, i.e.
Re(v, Dv)h ≤ αkvk2h ∀v for some α independent of h, t and v. Then
the energy estimate follows, since dkuk/dt = (ut , Du)h + (Du, ut )h =
2Re(u, Du)h .
For initial boundary value problems the Fourier method cannot be
used, the energy method on the other hand can be used. The calculation
become more difficult, due to boundary terms that do not cancel as in the
periodic case. They have to be estimated.
The most general technique for analyzing stability is the GKS method
named after the authors of [5]: Bertil Gustafsson, Heinz-Otto Kreiss and
Arne Sundström. The Laplace transform is applied in time and all but
one space variable are Fourier transformed. That leads to an eigenvalue
problem and the stability condition is that there are no eigenvalues with
real part larger than zero.
For fully discrete initial boundary value problems both the energy
method and the normal mode analysis also known as GKS method can
be used to show stability, but the calculations in general become more
difficult.

7
3 Paper A
In this paper high order difference schemes with the summation by parts
(SBP) property were studied [11].
To use a SBP operator to discretize an initial boundary value problem
(IVBP) the analytical boundary conditions must be imposed in a way
that does not destroy the SBP property.
Currently there are two methods described in the literature, the pro-
jection method and the simultaneous approximation term (SAT) method.
The theory of the projection method can be found in [10] and is out-
lined here. Olsson’s idea was that the boundary condition is fulfilled by
projecting the discrete solution of the initial value problem to the vector
space where the boundary condition is fulfilled. Using SBP operators he
could give an energy estimate for the semi-discrete case.
Let the boundary condition be written in the form

LT v = g (7)

were L is a rectangular matrix, v the vector of unknowns and g = g(t) a

known function. Let H be the diagonal positive definite matrix defining
the discrete scalar product (·, ·)h . Then the matrix

P = I − H−1 L(LT H−1 L)−1 LT (8)

where I is the identity matrix, defines a projection with the following

properties
• P2 = P
• HP = PT H
• v = Pv ⇔ LT v = 0.
Given the linear advection equation

ut + cux = 0 (9)

were c is the advection speed. The projection method for the semi discrete
problem becomes

vt + PcDv = (I − P)g̃t (10)

v(0) = f (11)

where g̃ = [g, 0, . . . , 0]T and f is the initial condition.

In the simultaneous approximation term (SAT) [2] method one does
not impose the exact boundary conditions which might destroy the SBP
property. Instead the boundary conditions are imposed as a penalty term
at the same accuracy as the discretization.
Below is the SAT formulation for (9).
τ −1
vt + Dv = − H e(LT v − g(t)) (12)
2h
v(0) = f (13)

where a bound on the parameter τ ≥ 1 is given by the energy method

and e = (1, 0, . . . , 0)T .
The two ways of imposing the boundary conditions were evaluated
on the 1D isentropic linearized Euler equations. No noticeable differences
were observed between the projection or SAT method, neither in accuracy

8
nor CPU time to solve the problem for the isentropic linearized Euler
equations.
A comparison between initial data of different regularity was also per-
formed, C 0 initial data and C ∞ initial data to be precise. As one may
suspect the C 0 initial data lead to a solution with high frequency oscilla-
tions, but the oscillations did not affect the order of accuracy. Howewer
for non linear problems artificial dissipation has to be used to damp the
oscillations which otherwise could grow exponentially [4].

4 Paper B
In aeroacustics Dispersion Relation Preserving schemes (DRP) [14] have
been given a lot of interest because the schemes have smaller dispersion
error than standard finite difference schemes. The price you have to pay
for that is to lower the formal accuracy of the method. Since the SBP
method studied on Paper A is proven stable the question arises: Can we
use DRP schemes instead of standard finite difference schemes in the SBP
operator? The answer is yes, and the paper deals with the derivation of
such difference operators.
Lets first explain the DRP schemes. If the formal accuracy of a fi-
nite difference method is lowered by two orders a free parameter is given
that can be chosen such that the wave number is approximated better.
The dimensionless
P wave number for the stencil −αl . . . − α1 0 α1 . . . αl is
hk̃ = lj=1 2αj sin(hk). The approximation of the wave number is good
for small wave numbers and gets worse for larger wave numbers. The free
parameter φ can be chosen such that DRP schemes gives a good approx-
imation for larger wave numbers than standard finite difference schemes.
For example one can minimize
Z π/2
E(φ) = |hk − hk̃(φ)|2 d(hk) (14)
−π/2
,
where k and k̃ are the exact and the approximate wave numbers, respec-
tively. Figure 2 shows the approximate wave number for the 2nd and
6th order standard centered difference schemes and also a 4th order DRP
scheme.

9
 3.5

2.5
Approximate wave number

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5
kh

Figure 2: Approximate non-dimensional wave number vs exact for 2nd order

standard centered difference method (SC2) (dotted), SC6 (dot-dashed), 4th
order DRP scheme (dashed).

Given a DRP scheme of order 2τ with coefficients αj , j = 1 . . . τ + 1

a summation by parts operator Q which is accurate of order τ near the
boundary and a diagonal norm matrix H with property that h(HQ +
(HQ)T ) = diag(−1, 0, . . . , 0, 1) will be derived.
For simplicity we approximate dx d
for x = [0, ∞) and let h = 1. The
difference operator Q has to be modified for rows 1 . . . 2τ .


q0,0 q0,1 ... q0,2τ −1
 .. ..
 q1,0 0 . . qτ −1,2τ 0 ...

 .. .. .. .. ..
 . . . q2τ −2,2τ −1 . . 0 ...
q2τ −1,0 ... q2τ −1,2τ −2 0 q2τ −1,2τ ... q2τ −1,3τ 0 ...

In order for HQ to be nearly antisymmetric and have the SBP property

the following relations must hold

h0,0 q0,0 = −1/2

hi,i qi,j = −hj,j qj,i 0 ≤ i < 2τ − 1 i < j ≤ 2τ − 1
hi,i qi,j = αj−i τ − 1 ≤ i < 2τ 2τ ≤ j ≤ τ + i + 1

Requiring that polynomials of degree τ and lower are exactly differen-

tiated by Q leads to τ + 1 systems of equations. That together with the

10
 conditions on HQ makes it possible to find closed expressions for Q and
H.
The summation by parts operators for DRP schemes were tested for
the 1D wave equation in first order form
µ ¶ µ ¶
1 0 uI
ut + ux = 0, u = ,
0 −1 uII
0 ≤ x ≤ 1, t ≥ 0,

With the initial data

uI (x, 0) = sin 2πx, uII (x, 0) = − sin 2πx, 0 ≤ x ≤ 1

and boundary conditions

uI (0, t) = uII (0, t), uII (1, t) = uI (1, t), t ≥ 0,

The methods used in the numerical experiment were the second, third
and fourth order DRPSBP methods, also denoted SBP-τ − 2τ (2(τ + 1)).
For the SBP-4-8(10) the free parameters were used to reduce the spectral
radius of the method and therefore making it possible to use larger time
steps. The parameters found reduced the spectral radius by about a factor
of 10.
The global order of accuracy is given in Table 1 and is in agreement
with theory, saying that the global accuracy is one order higher than the
accuracy at the boundaries if the order in the interior is at least one order
higher than the accuracy at the boundaries [3].

Table 1: Order of accuracy for uI using SBP 2-4(6), SBP 3-6(8) and SBP
4-8(10) for the 1D test case at t = 1.5.

# of grid points / SBP 2-4(6) SBP 3-6(8) SBP 4-8(10)

101
202 3.0137 3.9141 4.6758
401 3.0106 4.1361 4.6000
801 3.0083 4.3108 4.5678

The numerical experiments show that the DRPSBP method behaves

according to theory in the 1D case. When taking into account that the 4th
order Runge-Kutta was used in time, explaning the loss of accuracy for
the SBP-4-8(10) cf. last column in Table 1. The next step is to evaluate
the method on a 2D problem, which is done in the next paper, paper C.

5 Paper C
The 2D linearized Euler equations in conservative form [1] are used as a
model of sound propagation. Denoting density ρ, velocity in x-direction
u, in y-direction v, specific total energy E, specific total entalphy H and
pressure p. The equations are formulated in the variables ρ0 , (ρu)0 , (ρv)0
and (ρE)0 that are perturbations of a reference state ρ0 , ρ0 u0 , ρ0 v0 , ρ0 E0 .

11
    
ρ0 (ρu)0
∂  0 
 (ρu)0  + ∂
 ρ0 u0 u0 + (ρu)0 u0 + p0 
 +
∂t  (ρv)  ∂x  ρ0 v0 u0 + (ρv)0 u0 
(ρE)0 0
ρ0 H0 u + (ρH) u0 0
 
(ρv)0
∂  ρ0 u0 v 0 + (ρu)0 v0 
=0 (15)
∂y  ρ 0 0
0 v0 v + (ρv) v0 + p
0 
0 0
ρ0 H0 v + (ρH) v0

The acoustic quantities are defined by ρ0 = ρ − ρ0 , (ρu)0 = ρu − ρ0 u0 ,

(ρv)0 = ρv − ρ0 v0 and (ρE)0 = ρE − ρ0 E0 . With (ρH)0 = (ρE)0 + p0 .
The variables u0 , v 0 and p0 can be calculated using

(ρu)0 − ρ0 u0
u0 = (16)
ρ0
µ ¶
1 ¡ ¢
p0 = (γ − 1) (ρE)0 − 2 (ρu)0 · u0 − ρ0 u0 · u0 (17)
2

with u = (u, v)T .

The high order SBP operator based on the Tam and Webb DRP
scheme was implemented in a FORTRAN program [9] and tested for
an aeroacoustic test problem from [12]. In the domain [−100, 100] ×
[−100, 100] the mean velocities were chosen as u0 = 0.5, v0 = 0. The
initial conditions

x2 +y 2
p0 (x, y, 0) = e−log(2) 9 (18)
2 2 (x−67)2 +y 2
0 −log(2) x +y
9
−log(2) 25
ρ (x, y, 0) = e + 0.1e (19)
(x−67)2 +y 2
u0 (x, y, 0) = 0.04ye −log(2) 25 (20)
(x−67)2 +y 2
v 0 (x, y, 0) = −0.04(x − 67)e −log(2) 25 (21)

.
result in an acoustic pulse first located at the center and propagating
with the speed of sound in all directions transported by the mean veloc-
ity and entropy and vorticity waves propagating with the mean velocity
originating near the right boundary.
The results were compared with those obtained with a third order SBP
operator SBP-3-6 based on a sixth order standard central finite difference
stencil.
The 2D linearized Euler equations were solved using those SBP opera-
tors to discretize x and y derivatives and using the standard fourth order
Runge-Kutta method to discretize the time derivative.
The numerical results for the 2D linearized Euler equations show that
the better approximation of the wave number for the DRP based SBP
operator of fourth order (second order near the boundary) results in a
smaller L2 -error than for a SBP operator based on standard sixth order
(third order near the boundary) stencil. The results are valid when the
problem is not fully resolved, which usually is the case for large scale
applications.

12
 Application of a sixth order filter affects the good wave approximation
of the DRPSPB operator. That problem can be solved by using a filter or
artificial dissipation that apply less dissipation for wave numbers hk < π/2
than the present filter.
The development and implementation of accurate non reflecting bound-
ary conditions is future work.

References
[1] M. Billson, L.-E. Eriksson, and L. Davidson. Acoustic Source
Terms for the Linear Euler Equations on Conservative Form. In
8th AIAA/CAES Aeroacoustics Conference, Breckenridge, Colorado,
(2002). AIAA Paper 2002-2582.
[2] M.H. Carpenter, D. Gottlieb, and S. Abarbanel. Time-Stable Bound-
ary Conditions for Finite-Difference Schemes Solving Hyperbolic
Systems: Methodology and Application to High-Order Compact
Schemes. J. Comp. Phys., 111:220–236, (1994).
[3] B. Gustafsson. The Convergence Rate for Difference Approxima-
tions to Mixed Initial Boundary Value Problems. Math. Comp.,
29(130):396–406, (1975).
[4] B. Gustafsson, H.-O. Kreiss, and J. Oliger. Time Dependent Problems
and Difference Methods. John Wiley & Sons, New York, (1995).
[5] B. Gustafsson, H.-O. Kreiss, and A. Sundström. Stability Theory of
Difference Approximations for Mixed Initial Boundary Value Prob-
lems. II. Math. Comp., 26(119):649–686, (1972).
[6] B. Müller and S. Johansson. Strictly Stable High Order Difference
Approximations for Computational Aeroacoustics. In Computational
Aeroacoustics: From Acoustic Sources Modeling to Far-Field Radi-
ated Noise Prediction”, EUROMECH Colloquium 449, (2003).
[7] B. Müller and S. Johansson. Strictly Stable High Order Difference
Approximations for the Euler Equations. In Proceedings of the Tenth
International Congress on Sound and Vibration, (2003).
[8] B. Müller and S. Johansson. Strictly Stable High Order Difference
Approximations for Low Mach Number Aeroacoustics. In P. Neit-
taanmäki, T. Rossi, K. Majava, and O. Pironneau, editors, ECCO-
MAS 2004, (2004).
[9] C. L. Müller. High order accurate numerical solution of the lin-
earized Euler equations for sound propagation in the atmosphere.
Master’s thesis, Uppsala University, Department of Scientific Com-
puting, (2004).
[10] P. Olsson. Summation by Parts, Projections, and Stability. I. Math.
Comp., 64:1035–1065, (1995).
[11] B. Strand. Summation by Parts for Finite Difference Approximations
for d/dx. J. Comp. Phys., 110:47–67, (1994).
[12] C.K.W. Tam. Benchmark Problems and Solutions. In J.C. Hardin,
J.R. Ristorcelli, and C.K.W. Tam, editors, ICASE/LaRCWorkshop
on Benchmark Problems in Computational Aeroacoustics (CAA),
(1995). NASA CP 3300.
[13] C.K.W Tam. Computational aeroacoustics: Issues and methods.
AIAA J., 33:1788–1796, (1995).

13
[14] C.K.W Tam and J. C. Webb. Dispersion-Relation-Preserving Finite
Difference Schemes for Computational Acoustics. J. Comp. Phys.,
107:262–281, (1993).
[15] J. Westerlund. High order Simulation of Rocket Launch Noise. Mas-
ter’s thesis, Uppsala University, Department of Scientific Computing,
(2002).

14
Paper A
Paper B
Paper C
 NUMERICAL SOLUTION OF THE
LINEARIZED EULER EQUATIONS USING
HIGH ORDER FINITE DIFFERENCE
OPERATORS WITH THE SUMMATION BY
PARTS PROPERTY
Stefan Johansson

Department of Information Technology

Scientific Computing
Uppsala University
P.O. Box 337
SE 75105, Uppsala
Sweden
 Abstract
We have used high order finite difference methods with the summa-
tion by parts property (SBP) on the 1D linearized Euler equations. The
boundary conditions are imposed with both the projection method and
the simultaneous approximation term method (SAT) for comparison. The
formal fourth order of accuracy of the high order SBP operator was verified
with both the projection method and the SAT method. Some relatively
large errors were observed at the artificial boundaries and further investi-
gations are needed to improve the non-reflecting boundary conditions.
Key words: linearized Euler equations, computational aeroacoustics, high order
difference method, summation by parts, non-reflecting boundary conditions

1 Introduction
Most natural phenomena of interest are governed by partial differential equa-
tions (PDEs), yet no general technique is known to find the exact solution to
a given well-posed PDE. Therefore numerical methods have been used even
long before the birth of digital computers, see for example the work by L. F.
Richardson [9].
PDEs governing acoustics often require high order methods to reach accuracy
requirements: low dissipation and dispersion errors, cf. Tam [11].
To exemplify there are large scale disparities between the eddy scale l and
the acoustic wavelength λ ∼ lM −1 , where M is the Mach number. There are
also large energy density disparities between the hydrodynamic near field and
the acoustic far field cf. [4].
A technique that can be used to avoid resolving the different scales was
proposed by Lighthill [7] in the 1950’s where one first solves the flow field and
uses that solution as a source term to the wave equation to get the acoustic
field.
The linearized Euler equations have received interest since they can be used
to model refractional effects and reflections at solid boundaries cf. M. Billson
et al. [1]. Direct simulation of these phenonema has come into reach for our
computers only in recent years.
In the numerical experiments I have studied isentropic and non isentropic
sound propagation, governed by the linearized Euler equations in one space di-
mension. Discretization was done with a finite difference method, more precisely
with a high order summation by parts operator (SBP) [6][10]. The boundary
conditions were imposed with the simultaneous approximation term method
(SAT) introduced by Carpenter et al. [2] and in some cases by the projection
method described by Olsson [8]. The classical four step Runge-Kutta method
was used for time marching.
The most important advantage of the summation by parts operators is that
they lead to strictly stable high order finite difference methods.
The goal of this investigation was to check the accuracy of the summation
by parts operator for the linearized Euler equations and to compare the simul-
taneous approximation term method and the projection method both used to
impose boundary conditions.

2
2 The governing equations
The 1D nonlinear Euler equations read

ut + Aux = 0 (1)

where u = (ρ, u, p)T and

 
u ρ 0
A= 0 u 1/ρ  (2)
0 γp u

where u is the velocity, p is the pressure and ρ is the density. Time is denoted
by t, x is the space coordinate and γ is the ratio of specific heats, here γ = 1.4
for air. We then linearize around the mean values (R, U, P ). We replace u by
U + ²u0 and analogous for p and ρ, and neglect high order terms of O(²2 ) and
get
u0t + Bu0x = 0 (3)
where u0 = (ρ0 , u0 , p0 )T and
 
U R 0
B= 0 U 1/R  (4)
0 γP U

The eigenvalues of B are U − a, U and U + a, the corresponding eigenvectors

are the columns in the matrix
 
R/a 1 R/a
R =  −1 0 1  (5)
Ra 0 Ra

with inverse  
0 −1/2 1/(2Ra)
R−1 = 1 0 −1/a2  (6)
0 1/2 1/(2Ra)
where a = γP/R is the speed of sound. The linearized Euler equations are often
used to model sound propagation.
To simplify things we can assume that the change in entropy is zero, the so
called isentropic case. Then
µ ¶γ µ ¶ γ−1
γ
p ρ T
= = (7)
p∞ ρ∞ T∞

where the subscript ∞ denotes a reference state and T is temperature.

The 1D linearized Euler equations for the isentropic case read

u0t + Cu0x = 0 (8)

where µ ¶
U a
C= (9)
a U

3
and u0 = (u0 , ρ0 )T . U , a, u0 and ρ0 are the mean flow velocity, mean speed of
sound, velocity and density perturbations, respectively. In the following, the
prime 0 to indicate perturbation variables will be omitted i.e. from here on, we
use the notation u = u0 and ρ = ρ0 , if not stated otherwise. The eigenvalues of
A are U − a and U + a, and the corresponding eigenvectors are the columns in
µ ¶
1 1 1
R= √ . (10)
2 1 −1

Define w = R−1 u0 then we can multiply the above systems (3) and (8) from
the left by R−1 to get
wt + Λwx = 0 (11)
where Λ is a diagonal matrix containing the eigenvalues λi of B and C, respec-
tively.
The reason for transforming the system is to be able to use the SAT method
described in section 5 for implementing non-reflecting boundary conditions and
derive exact solutions.

3 Summation by parts operators

When doing numerical calculations we must establish an upper bound on the
growth of the solution. In the continuous case the energy method is often used,
for example, with the usual L2 scalar product and norm
Z b
(u, v) = uv dx , ||u|| = (u, u)1/2 . (12)
a

We have for the simplest hyperbolic equation, sometimes called the Kreiss equa-
tion,
ut = ux (13)
the following energy growth

1 d||u||2 1
= (u, ut ) = (u, ux ) = [u2 ]ba . (14)
2 dt 2
That is the energy growth in time is governed by the boundary values. In the
last step, we used integration by parts

(u, vx ) = [uv]ba − (ux , v) (15)

In the discrete case we want to find an operator D and a scalar product
Rb
H that approximates the derivative d/dx and the integral a with the same
properties as the continuous case.

(u, Dv)h = un vn − u0 v0 − (Du, v)h (16)

where
(u, v)h = huT Hv, where h is the step size. (17)
Since equation (16) is the discrete analogue of equation (15), it is called
summation by parts property.

4
 An example of such an operator D and scalar product H is
 
−1 1 0 ... 0  
 .. ..  0.5 0 ... 0
 −0.5 0 0.5 . .   0 1 0 
1   
. .. . .. . ..   .. . .. .. .. .. 
D=  0 0 ,H =  . . . . .
h   
 . .   0 1 0 
 .. . . −0.5 0 0.5 
0 ... 0 0.5
0 ... 0 −1 1

Operators if this kind were described by Kreiss and Scherer [6], and high order
operators were later constructed by Strand [10]. High order operators D have
the following structure: in the interior a standard centered high order finite dif-
ference stencil is used and near the boundary when the standard stencil cannot
be used, a small dense matrix is used instead. E.g. the boundary matrix giving
third order is 6 × 9. The stencil giving sixth order and that boundary matrix
giving third order are given in the Appendix. For diagonal norms H Strand [10]
constructed matrices Q giving order s = 1, . . . , 4 corresponding to stencils of
order 2s in the interior.

4 Projection method
To apply high order operators of the SBP type to an initial boundary value
problem (IVBP) the analytical boundary conditions must be satisfied in a cer-
tain way in order not to destroy the SBP property. Currently there are two
popular methods, the projection method and the simultaneous approximation
term (SAT) method. The latter method will be described in the next section.
The theory of the projection method can be found in [8] and is outlined here.
Olsson’s idea was that the boundary condition was fulfilled by projecting the
discrete solution to the initial value problem to the vector space where it is
fulfilled. Using SBP operators he could give an energy estimate for the semi-
discrete case.
Let the boundary condition be written in the form

LT v = g (18)

were L is a rectangular matrix, v the vector of unknowns and g = g(t) a known

function. Then the matrix

P = I − H−1 L(LT H−1 L)−1 LT (19)

where I is the identity matrix, defines a projection with the following properties
• P2 = P
• HP = PT H
• v = Pv ⇔ LT v = 0.
Given the linear advection equation

ut + cux = 0 (20)

5
were c is the advection speed. The projection method for the semi discrete
problem becomes

vt + PcDv = 0 (21)
v(0) = f (22)

where f is the initial condition. In general the right hand side in (21) involves
time derivatives of g, but in our case it is zero because we implement non-
reflecting boundary conditions.
Given two matrices Am,n and Bp,q . The Kronecker product ⊗ is defined as
 
a1,1 B . . . a1,n B
 .. .. 
A⊗B= . . .
am,1 B . . . am,n B
For systems the projection method can be expressed using the Kronecker prod-
uct

vt + (I ⊗ P)(IN,N ⊗ C)(D ⊗ I)v = 0 (23)

v(0) = f (24)

were IN,N is the identity matrix of dimension N × N , C is the diagonal matrix

diag[c, . . . , c]T the dimension of the identity matrix I is the same as the number
of equations (in our case we have two equations) in the system, and the values
in v and f are ordered vector wise, for example v = [u0 , ρ0 , u1 , ρ1 , . . . , uN , ρN ]T .

5 Simultaneous approximation term

In the SAT method one does not impose the exact boundary conditions (b.c.)
which might destroy the SBP property and can lead to time-instabilities i.e. the
solution can suddenly blow up if you calculate long enough in time. Instead the
boundary conditions are imposed as a penalty term at the same accuracy as the
discretization. Below is the SAT formulation for (3) and is analogous for (8).
We start with the semi-discrete form of (3).

vt + D ⊗ Bv = 0 (25)

and for systems the SAT method reads:

vt + (IN,N ⊗ B)(D ⊗ I)v = SAT (26)

were IN,N is the identity matrix of dimension N × N , n is the number of equa-

tions in the system, the dimension of the identity matrix I is n, and the values in
v are ordered vector wise, for example v = [u0 , ρ0 , u1 , ρ1 , . . . , uN , ρN ]T . where

 −h−1 + −1
00 R(Λ R v0 − Φ0 (t))h−1 , j=0,
SATj = 0, 0 < j < N, (27)
 −1
hN N R(Λ− R−1 vN − ΦN (t))h−1 , j=N,

h00 and hN N are the first and last elements of the norm matrix H and Λ± =
1
2 (Λ ± |Λ|), with |Λ| = diag(|λi |). Non-reflecting boundary conditions are im-
plemented by Φ0 (t) = ΦN (t) = 0

6
6 The classical Runge-Kutta method for ODE’s
Using a summation by parts operator in the system
ut + Aux = 0 (28)
gives a large system of ODE’s
vt = Pv (29)
d
were P is a discrete approximation of −A dx .
To solve this system we use the classical Runge-Kutta method [3], which is
fourth order accurate.
v (1) = vn (30)
1
v (2) = v n + ∆tPv (1) (31)
2
1
v (3) = v n + ∆tPv (2) (32)
2
v (4) = v n + ∆tPv (3) (33)
µ ¶
1 (1) 1 (2) 1 (3) 1 (4)
v n+1 = n
v + ∆t Pv + Pv + Pv + Pv . (34)
6 3 3 6

7 Numerical experiments
The test cases are now described in more detail. The equations used are the
linearized isentropic Euler equations(8) and the linearized Euler equations (3),
with a sine pulse in the isentropic case and a Gaussian in the non-isentropic
case.

7.1 Isentropic test case

The initial values given are
½
0 (sin x−0.4
0.2 π)
4
if 0.4 ≤ x ≤ 0.6
u (x, 0) =
0 otherwise
The initial density perturbation ρ0 is set to zero, and the Mach number M = U/a
is 0.5.
The solution to the acoustic velocity field and the acoustic density field there
are two waves traveling in opposite directions and at different speeds (U + a = 1
and U − a = −1/3) see figures 1 and 2. There are no apparent differences in the
two methods used to impose the boundary conditions. The convergence rate is
close to four when a wave has reached the boundary, as it should be since the
overall accuracy is at most one order higher than at the boundary [5]. The same
tests were done with the SBP operator of second order with similar results.
Different time steps were also used with the high order SBP operator, varying
from CF L = 0.1 and CF L = 0.99. If a larger CF L number was used the
convergence rate dropped somewhat but not below four when the waves were
inside the domain and then dropped to about four at the boundary.
Since we have a sixth order stencil in the interior we should expect sixth
order accuracy until a wave reaches the boundary, and one explanation may be
that the initial condition has a discontinuous derivative.

7
 Some wiggles were observed in the solution see section 7.3, these can be
damped using artificial dissipation. This is often necessary when using finite
difference methods with no inherent dissipation as with the present SBP oper-
ators.
The time instants were chosen such that for the first time both waves were
inside the computational domain, for the second time t = 0.75 one had reached
the boundary and in the last case the first wave had left the domain and the
second had reached the other boundary. In some test runs the computation
continued until T = 20 to check for long time stability. Some small wiggles were
observed in the solution due to the fact that no artificial dissipation was added
to the numerical scheme, and also due to the fact that the initial condition has
a discontinuous first derivative. In more realistic test cases artificial dissipation
has to be added to damp high frequency oscillations.
Results using the projection method can be seen in table 1. The convergence
rate is computed with respect to a grid with half as many points. The columns
give the results from three time instants 0.25, 0.75 and 1.5.

Table 1: Order of accuracy for SBP 3-6 using the projection method for the
isentropic Euler equations
# of grid points / Time 0.25 0.75 1.5
100
200 4.7253 4.0004 3.9834
400 4.4069 4.0140 4.0369
800 4.2376 4.0138 4.0453

Results using the simultaneous approximation term method can be seen in

table 2. The convergence rate is computed with respect to a grid with half as
many points. The columns give the results from three time instants 0.25, 0.75
and 1.5.

Table 2: Order of accuracy for SBP 3-6 using the simultaneous approximation
term method for the isentropic Euler equations
# of grid points / Time 0.25 0.75 1.5
100
200 4.7253 4.0004 3.9885
400 4.4069 4.0140 4.0779
800 4.2376 4.0138 4.0961

For both the projection and SAT method we observe an overall forth order
accuracy, as should be expected. More surprising is that the order of accuracy
is not close to six when no wave has reached the boundary. Later experiments
indicated that reason for the lower accuracy was caused by the non-smooth
initial condition, cf. section 7.3.

8
Figure 1: Acoustic velocity and error fields for the isentropic linearized 1D Euler
equations using 400 points in the x-direction.

Figure 2: Density perturbations and error fields for the isentropic linearized 1D
Euler equations using 400 points in the x-direction.

The large errors observed at the boundaries in figures 1 and 2 were unex-
pected, because one had expected an error at the boundary that was comparable
to the error inside the computational domain. Future investigation will deal with
this problem.

7.2 Non-isentropic test case

7.2.1 Left going acoustic wave
The initial values are now instead given by
2
u0 (x, 0) = αe(−β(x−0.5) )
(35)
p0 (x, 0) = −Rau0 (x, 0) (36)
ρ0 (x, 0) = −Ru0 (x, 0)/a (37)
where the parameters are chosen as α = 1 and β = 250. The Mach number is
chosen to be 0.5.

9
Figure 3: Acoustic velocity end error fields for the non isentropic linearized 1D
Euler equation using 400 points in the x-direction, respectively.

The wave in figure 3 is moving to the left and therefore corresponds to the
eigenvalue U − a. The only Riemann invariant that is non zero with the initial
values used is the one corresponding to the same eigenvalue. The characteristic
variables are given by multiplying the primitive variables i.e. ρ, u, and p with
the inverse of the eigenvector matrix from the left.

Table 3: Order of accuracy for SBP 3-6 using the simultaneous approximation
term method for the non-isentropic Euler equations, with left going acoustic
wave.
# of grid points /Time 0.25 0.75 1.5
100
200 5.8772 5.7965 3.9386
400 5.8238 5.7486 3.9602
800 5.6634 5.5458 3.9678

In table 3 the time instants 0.25, 0.75 and 1.5 were the same as for the
isentropic case for convenience. For the first two time instants the wave is still
inside the domain, and has left at the third. Now we observe about sixth order
of accuracy, when no information has reached the boundary as we should expect.

7.2.2 Right going acoustic wave

The initial values are now given by
2
u0 (x, 0) = αe(−β(x−0.5) )
(38)
p0 (x, 0) = Rau0 (0, x) (39)
ρ0 (x, 0) = Ru0 (x, 0)/a (40)

where the parameters are chosen as, α = 1 and β = 250. The Mach number is
chosen to be 0.5.

10
Figure 4: Acoustic velocity and error fields for the non isentropic linearized 1D
Euler equation using 400 points in the x-direction, respectively.

The wave in figure 4 is moving to the right and therefore corresponds to the
eigenvalue U + a. The only Riemann invariant that is non zero with the initial
values used is the one corresponding to the same eigenvalue.

Table 4: Order of accuracy for SBP 3-6 using the simultaneous approximation
term method for the non-isentropic Euler equations, with right going acoustic
wave.
# of grid points /Time 0.05 0.15 0.3
100
200 5.9284 5.9189 3.9354
400 5.9424 5.9353 3.9585
800 5.8441 5.8203 3.9667

The reason for using different time instants was the greater speed of the right
going wave. So in order to capture it inside the domain smaller time instants had
to be used. At the first two time instants 0.05 and 0.15 the wave is still inside
the domain and at the last instant 0.3 it has left the computational domain.
As before we compare the numerical solution with the exact solution and the
results are similar to the isentropic case. Except of course the observed close to
sixth order accuracy when the wave was still inside the domain cf. table 4.

7.2.3 Right going entropy wave

The initial values are now given by
u0 (x, 0) = 0 (41)
0
p (x, 0) = 0 (42)
0 (−β(x−0.5)2 )
ρ (x, 0) = αe (43)
(44)
where the parameters are chosen as α = 1 and β = 250. The Mach number
is chosen to be 0.5. For the above initial conditions only the density ρ0 is non
zero.

11
Figure 5: Acoustic density and error fields for the non isentropic linearized 1D
Euler equation using 400 points in the x-direction, respectively.

Table 5: Order of accuracy for SBP 3-6 using the simultaneous approximation
term method for the non-isentropic Euler equations, with right going entropy
wave.
# of grid points /Time 0.25 0.75 1.5
hline 100
200 5.9243 5.9148 4.0473
400 5.9324 5.9253 4.0852
800 5.8960 5.8899 4.0885

In the above experiment see figure 5 and table 5, we have chosen the initial
condition such that only one Riemann invariant, namely entropy, was non-zero.
All three test cases for the non-isentropic Euler equations displayed sixth
order accuracy and fourth order when the wave reaches the boundary as theory
predicts. Also the wiggles that were observed in the isentropic case with the non-
smooth initial condition were not observed and are discussed in the following
section.

7.3 Comparison between C 0 and C ∞ initial data

For the isentropic case the following C 0 initial data was used cf. figure 6
½
(sin x−0.4
0.2 π)
4
if 0.4 ≤ x ≤ 0.6
u0 (x, 0) =
0 otherwise.

and the initial density perturbation ρ0 is set to zero. As before the Mach number
M = U/a was chosen to be 0.5.

12
 1
Truncated sin4
Gaussian
0.9

0.8

0.7

0.6

0.5
y

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

Figure 6: The two initial conditions, solid for the truncated sin4 curve and
dashed for the Gaussian curve.

Due to the discontinuity in the first derivative there are wiggles in the so-
lution as can be seen in Fig. 7. The discontinuity leads to high frequency
oscillations in the Fourier domain that cannot be represented by the finite dif-
ference method.
These wiggles are not present when we use the C ∞ initial data that were
used in the non isentropic case
2
u0 (x, 0) = αe(−β(x−0.5) )

with α = 250 and β = 0.5.

13
 Time: 0.555 Time: 1.555
0.3 0.2
Truncated sin4
Gaussian
0.2
0.1

0.1
0

0
−0.1

−0.1
u

u
−0.2
Truncated sin4
−0.2 Gaussian

−0.3
−0.3

−0.4
−0.4

−0.5 −0.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x x

(a) The fastest acoustic wave is (b) The second, slower acoustic
leaving the domain, notice the wave is leaving, and again notice
wiggles at the right boundary. the wiggles at the left boundary
this time.

−3 Time: 2.155
x 10
6
Truncated sin4
Gaussian

0
u

−2

−4

−6
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x

(c) Both waves have left the domain, pre-

sented for completeness.

Figure 7: Numerical solution using different initial data, solid denotes the one
with truncated sin4 and dashed the Gaussian curve.

We have seen the high frequency oscillations that occur in the numerical
solution when using the truncated sin4 curve as initial condition. These oscil-
lations are most likely the cause of the fact that sixth order accuracy was not
observed when no wave had reached the boundary in the isentropic case.

7.4 Comparison between SBP 1-2 and SBP 3-6

The summation by parts operator SBP 3-6 has third order accuracy at the
boundary and therefore the overall accuracy is fourth order[5]. The SBP 1-2
operator therefore is of second order accuracy, the figures 8 and 9 show the error
compared to the exact solution.
The time to solve the non isentropic Euler equations using with SBP 3-6 is
only about 50 % more per time step than with SBP 1-2, and the figure 8 shows
the advantage of the high order method on this case.

14
 For the SBP 3-6, the error is about one order of magnitude smaller until the
wave reaches the boundary were the error has a jump. When the wave leaves
the boundary the accuracy is not sixth order as it was in the interior but fourth
order.
When using the SBP 1-2 it is also apparent that the error grows linearly
with time, this can also be seen when zooming in on 9. This can be observed
with all finite difference methods.

Error for density with 100 points using SBP 1−2

0.2

0.1

−0.1

−0.2
6

Time 2 1
0.8
0.6
0.4
0 0.2 x
0

Figure 8: The error with time for the density with 100 and 400 points using
SBP 1-2.

Error for density with 100 points using SBP 3−6

−3
x 10
4

−2

−4
6

2 1
0.8
0.6
0.4
Time 0 0.2
0
x

Figure 9: The error with time for the density with 100 and 400 points using
SBP 3-6.

7.4.1 Mach number dependence

In order to investigate the Mach number dependence on the numerical solution
a test case was chosen with Mach number equal to 0.1. The initial values are

15
now given by
2
u0 (x, 0) = αe(−β(x−0.5) )
(45)
2
p0 (x, 0) = −αe(−β(x−0.5) )
(46)
2
0 (−β(x−0.5) )
ρ (x, 0) = αe (47)
where the parameters are chosen as α = 1 and β = 250.

In the previous test cases for the non-isentropic Euler equations there are
now three waves present a left and a right going acoustic wave and a right going
entropy wave.
The same behavour is observed as for the three test cases for the linear non-
isentropic Euler equations with Mach number 0.5. That is that there are some
problems at the boundaries which were not anticipated. Other then that the
only difference is that the calculations must be run for longer time in order to
let the slowest wave leave the domain, namely the entropy wave.

8 Conclusions
The summation by parts operators of overall fourth order are used together
with the projection method and the simultaneous approximation term method
to implement the boundary conditions and the classical fourth order Runge-
Kutta method for time-stepping. The isentropic linearized 1D Euler equations
and the non isentropic linearized 1D Euler equations are considered. Note that
the projection method was not used for the non isentropic linearized Euler
equations.
No noticeable differences were observed between the projection or simulta-
neous approximation term method, neither in accuracy or time to solve the
problem for the isentropic linearized Euler equations.
A comparison between summation by parts operators with different orders
of accuracy showed that a 50 % increase in computation time for the high order
method reduced the error by about two orders magnitude compared with the
second-order method.
The errors of the reflected numerical waves are considerably larger then
the errors of the outgoing physical waves. These problems at the boundary

16
that were observed with the high order method were not anticipated. However,
when the same method and boundary conditions were implemented for the scalar
Kreiss equation, the same problem at the boundary was found. One possible
explanation of the problem is that the discrete method allows for unphysical
waves traveling in the wrong direction, which the boundary conditions will not
handle properly. Further investigations will be conducted to determine the
source of the problem.

9 Appendix
The SBP operator with sixth order accuracy in the interior and third near the
boundary (SBP 3-6) has a standard centered finite difference operator in the
interior of D

1 3 3
D(j, j − 3) = − , D(j, j − 2) = , D(j, j − 1) = − ,
60h 20h 4h
3 3
D(j, j) = 0, D(j, j + 1) = , D(j, j + 2) = − ,
4h 20h
1
D(j, j + 3) = for j = 7, . . . , N − 6
60h
H is a diagonal matrix with ones in the diagonal except for some elements
in the beginning and end of the diagonal. In this case the first ones read
13649/43200 12013/8640 2711/4320 5359/4320 7877/8640 and 43801/43200,
and in reverse order at the end.
Below is the boundary operator for D in the SBP 3-6 operator, namely
D(1:6,1:9) = Q/h. The indices refer to indices in D but the same operator is
located in the bottom right half corner with opposite sign and the elements are
in a different order as the MATLAB code shows.
D(N-5:N,N-8:N) = -Q(6:-1:1,9:-1:1)/h;

Q( 1 , 1 )= −1.5825335189391164188
Q( 1 , 2 )= 1.9968007424231323418
Q( 1 , 3 )= .47988863653014872884 ∗ 10−2
Q( 1 , 4 )= −.66986592424353432486
Q( 1 , 5 )= .25079981439421691455
Q( 1 , 6 )= 0
Q( 1 , 7 )= 0
Q( 1 , 8 )= 0
Q( 1 , 9 )= 0
Q( 2 , 1 )= −.45374732928216654180
Q( 2 , 2 )= 0
Q( 2 , 3 )= .20413995948833208469
Q( 2 , 4 )= .42505341435666916396
Q( 2 , 5 )= −.19379006076750187297
Q( 2 , 6 )= .18344016204667166126 ∗ 10−1
Q( 2 , 7 )= 0
Q( 2 , 8 )= 0
Q( 2 , 9 )= 0
Q( 3 , 1 )= −.24160826263371449650 ∗ 10−2

17
Q( 3 , 2 )= −.45229312676749047092
Q( 3 , 3 )= 0
Q( 3 , 4 )= .23791958686831427518
Q( 3 , 5 )= .34541374646501905816
Q( 3 , 6 )= −.12862412393950571745
Q( 3 , 7 )= 0
Q( 3 , 8 )= 0
Q( 3 , 9 )= 0
Q( 4 , 1 )= .17061018846799776078
Q( 4 , 2 )= −.47641039995023947254
Q( 4 , 3 )= −.12035827579772345587
Q( 4 , 4 )= 0
Q( 4 , 5 )= .42710082726876904895
Q( 4 , 6 )= −.14377682403433476395 ∗ 10−1
Q( 4 , 7 )= .13435342414629595074 ∗ 10−1
Q( 4 , 8 )= 0
Q( 4 , 9 )= 0
Q( 5 , 1 )= −.86915492361728238331 ∗ 10−1
Q( 5 , 2 )= .29554398882823409928
Q( 5 , 3 )= −.23775972239854428505
Q( 5 , 4 )= −.58114341331302103170
Q( 5 , 5 )= 0
Q( 5 , 6 )= .75652321103635055647
Q( 5 , 7 )= −.16452964326520248826
Q( 5 , 8 )= .18281071473911387584 ∗ 10−1
Q( 5 , 9 )= 0
Q( 6 , 1 )= 0
Q( 6 , 2 )= −.25155437851495019140 ∗ 10−1
Q( 6 , 3 )= .79610054564964270222 ∗ 10−1
Q( 6 , 4 )= .17590922581676217438 ∗ 10−1
Q( 6 , 5 )= −.68025083141176381057
Q( 6 , 6 )= 0
Q( 6 , 7 )= .73970913906075203762
Q( 6 , 8 )= −.14794182781215040752
Q( 6 , 9 )= .16437980868016711947 ∗ 10−1

18
References
[1] M. Billson, L.-E. Eriksson, and L. Davidson. Acoustics Source Terms for
the Linear Euler Equations on Conservative Form. AIAA Paper, 2002-2582,
2002.
[2] M.H. Carpenter, D. Gottlieb, and S. Abarbanel. Time-Stable Bound-
ary Conditions for Finite-Difference Schemes Solving Hyperbolic Systems:
Methodology and Application to High-Order Compact Schemes. J. Comp.
Phys., 111:220–236, (1994).
[3] G. Dahlquist and Å. Björck. Numerical Methods. Prentice-Hall, 1974.
[4] Crighton D.G. Computational Aeroacoustics For Low Mach Number Flows.
In J.C. Hardin and M.Y. Hussaini, editors, Computational Aeroacoustics,
ICASE/NASA LaRC Series, 1995.
[5] B. Gustafsson. The Convergence Rate for Difference Approximations to
Mixed Initial Boundary Value Problems. Math. Comp., 29(130):396–406,
(1975).
[6] H.-O. Kreiss and G. Scherer. Finite Element and Finite Difference Meth-
ods for Hyperbolic Partial Differential Equations. In C. De Boor, editor,
Mathematical Aspects of Finite Elements in Partial Differential Equations.,
pages 195–211. Academic Press, Inc., (1974).
[7] J. Lighthill. On sound generated aerodynamically. I. General theory. Proc.
of the Royal Soc. of London A, 211:564–587, (1952).
[8] P. Olsson. Summation by Parts, Projections, and Stability. I. Math. Comp.,
64:1035–1065, (1995).
[9] L.F. Richardson. The Approximate Arithmetical Solution by Finite Dif-
ferences of Physical Problems involving Differential Equations, with an
Application to the Stresses in a Masonry Dam. Philo. Trans. of the Royal
Soc. of London, 210:307–357, (1910).
[10] B. Strand. Summation by Parts for Finite Difference Approximations for
d/dx. J. Comp. Phys., 110:47–67, (1994).
[11] C.K.W Tam. Computational aeroacoustics: Issues and methods. AIAA J.,
33:1788–1796, (1995).

19
Paper B
 HIGH ORDER FINITE DIFFERENCE OPERATORS WITH
THE SUMMATION BY PARTS PROPERTY BASED ON DRP
SCHEMES

Stefan Johansson
Division of Scientific Computing
Department of Information Technology
Uppsala University
P.O. Box 337
SE 75105 Uppsala
Sweden

Abstract.
Strictly stable high order finite difference methods based on Tam and Webb’s dis-
persion relation preserving schemes have been constructed. The methods have been
implemented for a 1D hyperbolic test problem, and the theoretical order of accuracy
is observed.

1 Introduction
Computational aeroacoustics (CAA) has been given increased interest because of
the need to better control noise levels from aircrafts, trains, cars, etc. due to in-
creased transport and stricter regulations from authorities. Other applications range
from simulating sound propagation in the atmosphere to improved design of musical
instruments.
Much of the current effort in CAA involves the development of schemes for approx-
imating derivatives in a way that better preserves the physics of wave propagation, a
phenomenon of less significance in typical aerodynamic computations. An example of
such a scheme is the Dispersion Preserving Relation (DRP) scheme proposed by Tam
and Webb [5].
In scientific computing it is imperative that the numerical methods are stable. A
good approximation of the wave propagation requires high order methods and for
difference methods problems arise when applying boundary conditions. To accurately
prescribe boundary conditions the simultaneous approximation term (SAT) method [1]
can be used if the space derivatives are discretized by a summation by parts (SBP)
operator proposed by Kreiss and Scherer [3].
In this paper SBP operators are derived for DRP type schemes, which together with
the SAT method lead to strictly stable methods.

2 Theory
In recent years, central difference methods of the type proposed by Tam and Webb [5]
called Dispersion Preserving Relation schemes (DRP) have attracted interest in aeroa-
coustics. The formal accuracy of the method is lowered to get a better approximation
of the wave number.
2

Another development in the theory of finite difference methods is high order opera-
tors Q = h1 H −1 B with the summation by parts property (SBP) developed by Strand
[4], i.e. B + B T = diag(−1, 0, . . . , 0, 1), with a discrete norm H and step size h. A
consequence is that a stability proof done by the energy method for the continuous
problem is valid for the semi-discrete problem if a SBP operators are used.
What we have done is to construct a SBP operator for difference schemes of DRP
type. The motivation has been to combine the good wave resolution of DRP schemes
with the good stability properties of SBP operators when using using SAT to prescribe
boundary conditions in a strictly stable way, first shown by Carpenter et al. [1].
Consider the discrete function v approximating the continuous function u on {xj }N 0
where xj = hj, h = 1/N and let vj denote v(xj ). A classical central finite difference
method approximating du/dx at xm is
l
1X
(Qv)j = αj (vm+j − vm−j ) (2.1)
h j=1
where αk , k = 1 . . . l are chosen such that the accuracy is of order 2l.
Tam and Webb [5] proposed that the accuracy is lowered to 2(l − 1) leaving a free
parameter and then minimize the wave number error
Z π/2
|hk − hk̃|2 d(hk) (2.2)
−π/2

where k and k̃ are the exact and the approximate wave numbers, respectively.
A comparison between standard centered finite difference methods and DRP schemes
is given in figure 2.1. Noticeable is that the fourth order DRP scheme derived from
a sixth order standard scheme approximates the wave number about as well as the
eighth order standard scheme. In figure 2.2 it can be seen that the DRP schemes in an
interval near kh = 1 lies strictly above the line defining the exact group velocity. The
approximation of the wave number and group velocity can be improved for hk < 2 by
choosing a larger interval to optimize over in equation 2.2, at the expense of a poorer
approximation for kh > 2.
 3

3.5

2.5
approximate wave number

2
g
f
d e
1.5
c

b
1
a

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5
kh

Figure 2.1: Approximate wave number vs exact for 2nd order standard centered
difference method (SC2) (a) (dotted), SC4 (b), SC6 (c), SC8 (d) (dashed) and
DRP schemes with order four (e) six (f) and eight (g) (solid).
4

1.5

0.5

0
Approximate group velocity

−0.5

−1 a

−1.5
b

−2
c

−2.5 d
e

f
−3

g
−3.5
0 0.5 1 1.5 2 2.5 3 3.5
kh

Figure 2.2: Approximate group velocity vs exact for 2nd order standard centered
difference method (SC2) (a) (dotted), SC4 (b), SC6 (c), SC8 (d) (dashed) and
DRP schemes with order four (e) six (f) and eight (g) (solid).

3 Summation by Parts Operators

When doing numerical calculations, we must establish an upper bound on the growth
of the solution. In the continuous case the energy method is often used,
for example, with the usual L2 scalar product and norm
Z 1
(u, v) = uv dx , ||u|| = (u, u)1/2 . (3.1)
0

We have for the simplest hyperbolic equation, sometimes called the Kreiss equation,

 ut = ux , x ∈ [0, 1], t ≥ 0
u(x, 0) = f (x) , x ∈ [0, 1] (3.2)
u(1, t) = g(t) t ≥ 0


the following energy growth

1
1 d||u||2
Z
1 21 1
= (u, ut ) = (u, ux ) = uux dx = [u ]0 = (g(t)2 − u(0, t)2 ) (3.3)
2 dt 0 2 2

If g ≡ 0 it follows that kuk ≤ kf k and the problem is well posed.

In the discrete case we want to find an operator Q that in a modified norm, deter-
mined by H, gives the same estimate as the continuous case. That is Q must satisfy

(u, Qv)h = un vn − u0 v0 − (Qu, v)h (3.4)

 5

where
(u, v)h = huT Hv. (3.5)
duj
The matrix Q is a banded matrix were row j approximate . Of course Q has to be
dx
modified at the l first and l last rows, where l is equal to or larger than the bandwidth
of Q. The matrix H is a modified identity matrix, its l first elements are modified.
Since equation (3.4) is the discrete analogue of integration by parts, it is called the
summation by parts property.
An example of such an operator Q and scalar product H is
 
−1 1 0 ... 0
0.5 0 . . . 0
 
 .. .. 
 −0.5 0
 0.5 . . 
  0 1 0 
1  . .
 
Q=  0 . .. . .. . ..

, H = . . . . . . . .

.
0   . . . . . 
 
h


 .. .. 0 1 0
  

 . . −0.5 0 0.5  0 . . . 0 0.5
0 ... 0 −1 1

Operators if this kind were described by Kreiss and Scherer [3], and high order
operators were later constructed by Strand [4]. High order operators Q have the
following structure: in the interior a centered high order finite difference stencil is
used and near the boundary where the stencil cannot be used, a small dense matrix is
used instead.
Given a DRP scheme of order 2τ with coefficients αj , j = 1 . . . τ + 1 a summation by
parts operator Q which is accurate of order τ near the boundary and a diagonal norm
matrix H with property that h(HQ + (HQ)T ) = diag(−1, 0, . . . , 0, 1) will be derived.
d
For simplicity we approximate dx for x = [0, ∞) and let h = 1. The difference
operator Q has to be modified for rows 1 . . . 2τ .

q0,0 q0,1 ... q0,2τ −1


 .. ..

 q1,0 0 . . qτ −1,2τ 0 ...
 .. .. .. .. ..
 . . . q2τ −2,2τ −1 . . 0 ...
q2τ −1,0 ... q2τ −1,2τ −2 0 q2τ −1,2τ ... q2τ −1,3τ 0 ...

In order for HQ to be nearly antisymmetric and have the SBP property the following
relation must hold

h0,0 q0,0 = −1/2

hi,i qi,j = −hj,j qj,i 0 ≤ i < 2τ − 1 i < j ≤ 2τ − 1
hi,i qi,j = αj−i τ − 1 ≤ i < 2τ 2τ ≤ j ≤ τ + i + 1
and τ th order accuracy near the boundary leads to the following system of equations,
j = 0, ..., τ . Note that qi,i = 0 for i 6= 0

0j q0,0 + ... αj q0,α = j ∗ 0j−1

.. .. .. ..
. . . .
0j qβ,0 + ... αj qβ,α = j ∗ β j−1
.. .. .. .. ..
. . . . .
0j qγ,0 + ... γ j qγ,γ ... + (3τ )j qγ,3τ = j ∗ γ j−1
6

where α = max{2, 2τ − 1}, β = max{0, τ − 2}, γ = 2τ − 1 and j = 0, ..., τ , note that

qi,i = 0 for i 6= 0. The accuracy conditions come from requiring that polynomials of
degree τ and lower are exactly differentiated.
From these systems of equations, a system of equations with hi,i qi,j , j > i, i =
0, . . . 2(τ − 1) as unknowns can be derived. The solution is used to compute hi,i , i =
0, . . . , 2τ − 1, which is used to compute qi,j , i = 0, . . . 2τ − 1, j = 0, . . . , 3τ .
To illustrate the derivation of H and Q we let τ = 1. Then the following conditions
yield the SBP property.

= − 12


 h0,0 q0,0
 h0,0 q0,1 +h1,1 q1,0 = 0


h0,0 q0,2 −α2 =0 (3.6)
 h1,1 q1,2 −α1 =0



h1,1 q1,3 −α2 =0


First order accuracy near the boundary leads to two equation systems
½
q0,0 + q0,1 + q0,2 = 0
(3.7)
q1,0 + q1,2 + q1,3 = 0
½
q0,1 + 2q0,2 = 1
(3.8)
2q1,2 + 3q1,3 = 1
Using the conditions for antisymmetry of HQ and multiplying the first equation in
each system by h0,0 and the second equation by h1,1 yields,

− 21 + h0,0 q0,1 + h0,0 q0,2 = 0

½
(3.9)
−h0,0 q0,1 + α1 + α2 = 0
½
h0,0 q0,1 + 2h0,0 q0,2 = h0,0
(3.10)
2α1 + 3α2 = h1,1
Using that α1 = 1/2 − 2α2 , which is equivalent to requiring that the interior scheme
is second order accurate, the solution can be computed.
1
h0,0 = + α2 , h1,1 = 1 − α2 (3.11)
2
1 1/2−α2 α2
q0,0 = − 1+2α q0,1 = h0,0
q0,2 = h0,0
2
α2 −1/2 α1 α2 (3.12)
q1,0 = h1,1
q1,1 = 0 q1,2 = h1,1
q1,3 = h1,1

For τ > 2 the system of equations for HQ is undetermined thus leading to free param-
eters, one for τ = 3 and three for τ = 4. The free parameters were initially chosen such
that the bandwidth of Q was minimized, but as is seen in figure 3.1 the spectral radius
for the fourth order method is very large and leads to a restrictive CFL condition.
The spectral radius of Q can be made smaller if the three parameters are chosen as
[0.502, −0.1, 0.799], which are close to the values that minimize the bandwidth. These
values were found by trail and error.
The eigenvalues in figure 3.1 were computed using MATLAB, and a computation in
exact arithmetic using MAPLE showed that all eigenvalues are imaginary for SBP-2-
4(6). Computations for the other SBP operators were unfortunately too time consum-
ing to complete.
The original operators by Strand [4] are denoted by SBP-τ -2τ and the new ones
SBP-τ -2τ -(σ). The number in parenthesis denote the order of the difference method
 7

modified for better wave approximation, in this article σ = 2(τ + 1). The operators
that have been derived are SBP-1-2(4), SBP-2-4(6) SBP-3-6(8) and SBP-4-8(10), where
Tam and Webb’s DRP scheme corresponds to SBP-2-4(6) in the interior. They can be
found in the appendix.

Spectrum of Q for SBP−2−4(6) Spectrum of Q for SBP−3−6(8)

0.015 0.02

0.015
0.01

0.01

0.005
0.005

Im
Im

0 0

−0.005
−0.005

−0.01

−0.01
−0.015

−0.015 −0.02
−1 −0.5 0 0.5 1 1.5 −8 −6 −4 −2 0 2 4 6 8
Re x 10
−7 Re −4
x 10

Spectrum of Q for SBP−4−8(10) Spectrum of Q for modified SBP−4−8(10)

0.2 0.025

0.02
0.15

0.015
0.1
0.01

0.05
0.005
Im

Im

0 0

−0.005
−0.05

−0.01
−0.1
−0.015

−0.15
−0.02

−0.2 −0.025
−1.5 −1 −0.5 0 0.5 1 1.5 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Re −4
x 10 Re −4
x 10

Figure 3.1: Spectrum for SBP-2-4(6), SBP-3-6(8), SBP-4-8(10) and a modified

SBP-4-8(10), h = 1/100.

4 Numerical experiments
The summation by parts operators for DRP schemes were tested for the hyperbolic
system

uI
µ ¶ µ ¶
1 0
ut + ux = 0, u = ,
0 −1 uII
0 ≤ x ≤ 1, t ≥ 0,

With the initial data

uI (x, 0) = sin 2πx, uII (x, 0) = − sin 2πx, 0 ≤ x ≤ 1,

8

and boundary conditions

uI (0, t) = uII (0, t), uII (1, t) = uI (1, t), t ≥ 0,

the exact solution is

uI = sin 2π(x − t), uII = − sin 2π(x + t) .

The boundary conditions were imposed using the simultaneous approximation term
(SAT) method [1].
In the SAT method one does not impose the exact boundary conditions (b.c.) which
might destroy the SBP property and therefore strict stability. Instead the boundary
conditions are imposed as a penalty term at the same accuracy as the discretization.
In our case, the SAT formulation reads
µ ¶ µ ¶
1 0 1 1 0
vt + Q ⊗ v = − H −1 ⊗ P (4.1)
0 −1 h 0 1
where v = (v0I , v0II , . . . , vN
I II T
, vN ) denotes the numerical approximation of the exact
solution u, and the matrix P which is imposing the boundary conditions and reads

1 −1 0 ... 0 0
 
 0 0 ... 0 0 
.. .. .. ..
 
P = (4.2)
 
 . . . . 

 0 0 ... 0 0 
0 0 ... 0 −1 1

Table 4.1: Order of accuracy for uI using SBP 2-4 and SBP 2-4(6) for the 1D
test case at t = 1.5.

# of grid points / SBP 2-4 SBP 2-4(6)

101
202 3.0134 3.0137
401 3.0091 3.0106
801 3.0068 3.0083

Table 4.2: Order of accuracy for uI using SBP 3-6(8) for the 1D test case at
t = 1.5.

# of grid points / SBP 3-6(8)

101
202 3.9141
401 4.1361
801 4.3108
 9

Table 4.3: Order of accuracy for uI using SBP 4-8(10) for the 1D test case at
t = 1.5.

# of grid points / SBP 4-8(10)

101
202 4.6758
401 4.6000
801 4.5678

The orders of accuracy for SBP-2-4 (SBP operator, 2nd order near boundaries, 4th
order standard scheme in interior) and SBP-2-4(6) (SBP operator, 2nd order near
boundaries, 4th order DRP scheme derived from 6th order standard scheme in interior)
agree with theory, saying that the global order of accuracy for a finite difference scheme
is at most one order higher than the accuracy at the boundaries [2].

5 Conclusions
New strictly stable high order finite difference methods have been developed. They
are based on dispersion relation preserving type schemes proposed Tam and Webb [5]
and strictl stability is guaranteed by requiring that the methods have the summation
by parts property introduced by Kreiss and Scherer [3].
The methods have been developed using MAPLE and implemented in MATLAB.
Numerical experiments show that the order of accuracy corresponds to theory.
10

6 Appendix
6.1 SBP operators with diagonal norm and minimal bandwidth
6.1.1 Second order at the boundary, SBP-2-4(6)
A SBP operator with a fourth order DRP finite difference operator in the interior
and second order near the boundary (SBP 2-4(6)) has the following centered finite
difference operator in the interior of hQ

hQ(j, j − 3) = −α3 , hQ(j, j − 2) = −α2 , hQ(j, j − 1) = −α1 ,

hQ(j, j) = 0,
hQ(j, j + 1) = α1 , hQ(j, j + 2) = α2 , hQ(j, j + 3) = α3 for j = 7, . . . , N − 6
where αj has the following values
α1 = 0.79926642697415587 α2 = −0.18941314157932453 α3 = 0.026519952061497799
H is a diagonal matrix with ones in the diagonal except for the first and last four
elements of the diagonal. In this case the first four elements read

h0,0 = 0.34532668264616756, h1,1 = 1.2556866187281647,

h2,2 = 0.86931338127183581, h3,3 = 1.0296733173538330
and in reverse order for hn−3,n−3 , . . . , hn,n .
Below is the boundary operator for hQ for the SBP 2-4(6) operator. The same op-
erator is located in the bottom right half corner with opposite sign and the elements
are in a different order as this MATLAB code indicates.
Q(N-3:N,N-6:N) = -Q(3:-1:0,6:-1:0);

q00 = -1.4479043326991200 q10 = -0.50703996036005625

q01 = 1.8437129980973616 q11 = 0
q02 = -0.34371299809736176 q12 = 0.54223976216033715
q03 = -0.052095667300879370 q13 = -0.056319682880449493
q04 = 0 q14 = 0.021119881080168560

q20 = 0.13653680246090086 q30 = 0.017471584109301608

q21 = -0.78324253158386005 q31 = 0.068681853722050198
q22 = 0 q32 = -0.70418755285702870
q23 = 0.83408716489415180 q33 = 0
q24 = -0.21788821575736760 q34 = 0.77623301828214610
q25 = 0.030506779986175047 q35 = -0.18395459840223802
q26 = 0 q36 = 0.025755695145768824

6.1.2 Third order at the boundary, SBP-3-6(8)

A SBP operator with a sixth order DRP finite difference operator in the interior and
third order near the boundary (SBP 3-6(8)) has the following centered finite difference
operator in the interior of hQ

hQ(j, j−4) = −α4 , hQ(j, j−3) = −α3 , hQ(j, j−2) = −α2 , hQ(j, j−1) = −α1 ,
hQ(j, j) = 0,
hQ(j, j+1) = α1 , hQ(j, j+2) = α2 , hQ(j, j+3) = α3 hQ(j, j+4) = α4 for j = 9, . . . , N −8
 11

where αj has the following values

α1 = 0.8331572598964345 α2 = −0.2331572598964345
α3 = 0.05230549233656718 α4 = −0.005939804278316752

H is a diagonal matrix with ones in the diagonal except for the first and last six
elements of the diagonal. In this case the first six elements read

h0,0 = 0.3153550936462424, h1,1 = 1.393363420657677,

h2,2 = 0.6216064920179795, h3,3 = 1.246449063537576,
h4,4 = 0.9087199126756564, h5,5 = 1.014506017464869

and in reverse order for hn−5,n−5 , . . . , hn,n .

Below is the boundary operator for hQ for the SBP 3-6(8) operator. The same op-
erator is located in the bottom right half corner with opposite sign and the elements
are in a different order as this MATLAB code indicates.
Q(N-5:N,N-9:N) = -Q(5:-1:0,9:-1:0);

q00 = -1.585514266533103 q01 = 2.008723732799078

q10 = -0.4546274514421262 q11 = 0
q20 = 0.006638621722402879 q21 = -0.4773187801117445
q30 = 0.1664613995014180 q31 = -0.4554757153982624
q40 = -0.08600120227144731 q14 = 0.2769005367876363
q50 = 0 q15 = -0.02035922973188162

q02 = -0.0130855991986174834833673814004 q03 = -0.657942933867588344344421746018

q12 = 0.212941181087929195929055954261 q13 = 0.407450971157474941475221424938
q22 = 0 q23 = 0.307082033336054068764476763891
q32 = -0.153142387513228968467498042145 q33 = 0
q42 = -0.172627548513855444351101578662 q43 = -0.660390739163292203649761743354
q52 = 0.0485326149573270703688045031784 q53 = 0.0939867209644063450256722219284
q04 = 0.247819066800230419419438769975 q05 = 0
q14 = -0.180588228368106206106416068749 q15 = 0.0148235275648283216283776183329
q24 = 0.252362375273239592208687307356 q25 = -0.0792086803303361725384538561848
q34 = 0.481455867214575438317175291842 q35 = -0.0764973850672778007026589857013
q44 = 0 q45 = 0.847673568210825495387209342755
q54 = -0.759283668722722050186941866401 q55 = 0
q06 = 0 q07 = 0
q16 = 0 q17 = 0
q26 = -0.00955556988961586894406770497530 q27 = 0
q36 = 0.0419636019366228644454343113704 q37 = -0.00476538067384707719478469156032
q46 = -0.256577694231350970143666364671 q47 = 0.0575595313880133315974641085544
q56 = 0.821244276084627463843621091279 q57 = -0.229823437103968213426853599491
q08 = 0 q09 = 0
q18 = 0 q19 = 0
q28 = 0 q29 = 0
q38 = 0 q39 = 0
q48 = -0.00653645220651922535583776530795 q49 = 0
q58 = 0.0515575969349816747689080909099 q59 = -0.00585487338277167148733298046745
12

6.1.3 Fourth order at the boundary, SBP-4-8(10)

A SBP operator with a eight order DRP finite difference operator in the interior
and second order near the boundary (SBP 4-8(10)) has the following centered finite
difference operator in the interior of hQ

hQ(j, j − 5) = −α5 , hQ(j, j − 4) = −α4 , hQ(j, j − 3) = −α3 ,

hQ(j, j − 2) = −α2 , hQ(j, j − 1) = −α1 , hQ(j, j) = 0,
hQ(j, j + 1) = α1 , hQ(j, j + 2) = α2 , hQ(j, j + 3) = α3
hQ(j, j + 4) = α4 , hQ(j, j + 5) = α5 for j = 9, . . . , N − 8

where αj has the following values

α1 = 0.85710439841851208608 α2 = −0.26526216962115666981
α3 = 0.074805208507138722005 α4 = −0.014448456841621349730
α5 = 0.0013596285337740972877

H is a diagonal matrix with ones in the diagonal except for the first and last eight
elements of the diagonal. In this case the first eight elements read

h0,0 = 0.294851829648342276, h1,1 = 1.52599254960446488,

h2,2 = 0.25663709986386517, h3,3 = 1.79947333003289182,
h4,4 = 0.411348429226366286, h5,5 = 1.2793004001361369051,
h6,6 = 0.9230236540992415309, h7,7 = 1.009372707388694828

and in reverse order for hn−7,n−7 , . . . , hn,n .

Below is the boundary operator for hQ for the SBP 4-8(10) operator. The same
operator is located in the bottom right half corner with opposite sign and the elements
are in a different order as this MATLAB code indicates.
Q(N-7:N,N-12:N) = -Q(7:-1:0,12:-1:0);

q00 = -1.6957669911573197987 q01 = 2.0621682891199350282

q10 = -0.39845154764838537791 q11 = 0
q20 = -1.0060547060384519478 q21 = 2.6156550656068261280
q30 = 0.41657226346831719539 q31 = -1.6072314212261918344
q40 = -1.2098265916012883296 q41 = 5.8230938981601276548
q50 = 0.089325888656450236921 q51 = -0.72026816640720284437
q60 = 0 q61 = 0.15057109202244840663
q70 = 0 q71 = 0
q02 = 0.87566342176012784417 q03 = -2.5423300884267936392
q12 = -0.43989345194087425217 q13 = 1.8952714273978855804
q22 = 0 q23 = -10.439577939423169752
q32 = 1.4888706386814724978 q33 = 0
q42 = -10.496233449667786180 q43 = 7.2472876183287816283
q52 = 2.0946589947956824426 q53 = -2.7302904871947353579
q62 = -0.70025901086696570588 q63 = 1.2248064281916752937
q72 = 0.017998153763907768034 q73 = -0.080090074114033601879
 13

q04 = 1.6878317108800632472 q05 = - 0.38756634217601260433

q14 = - 1.5696803558226220357 q15 = 0.60382952310539918663
q24 = 16.823791823568735151 q25 = -10.441584991462022668
q34 = - 1.6566849467542517715 q35 = 1.9410466687451647647
q44 = 0 q45 = - 2.9362325326725905861
q54 = 0.94412120877136880657 q55 = 0
q64 = - 0.83167376571312553306 q65 = - 0.48635033688049237125
q74 = 0.10696008003992586465 q75 = 0.097204485324945230782
q06 = 0 q07 = 0
q16 = -0.091075595091402694223 q17 = -0.070788070790662511317
q26 = 2.5185588185387488660 q27 = 0.044924664119343000785
q36 = -0.62825343729493118302 q37 = -0.26246018679458559520
q46 = 1.8661905666948906603 q47 = -0.076694695407211351188
q56 = 0.35090496733375371850 q57 = 0.86342625760635763892
q66 = 0 q67 = 0
q76 = -0.78956252086787838706 q77 = 0
q08 = 0 q09 = 0
q18 = 0 q19 = 0
q28 = 0 q29 = 0
q38 = 0.00075557026107702591326 q39 = 0
q48 = 0.0033052965252138828252 q49 = -0.035124618972764619908
q58 = 0.0010627906734254224064 q59 = -0.011294029799477757896
q68 = 0.0014730159164782497904 q69 = -0.015653398238988014629
q78 = 0.0013470034644502469271 q79 = -0.014314293160353361934
q010 = 0 q011 = 0
q110 = 0 q111 = 0
q210 = 0 q211 = 0
q310 = 0 q311 = 0
q410 = 0 q411 = 0
q510 = 0.058473528577946443849 q511 = 0
q610 = 0.081043652754640918266 q611 = -0.28738393479202890305
q710 = 0.074110591617504789877 q711 = -0.26279903119968949743
q012 = 0
q112 = 0
q212 = 0
q312 = 0
q412 = 0
q512 = 0
q612 = 0
q712 = 0.84914560513122095062
14

6.2 Modified SBP-4-8(10) operator with diagonal norm and smaller spectral
radius
q00 = -1.695766991157320 q01 = 2.291032605468909
q10 = -0.4426726432456110 q11 = 0
q20 = 0.5330534182148309 q21 = -2.407040818496168
q30 = -0.1184427737693269 q31 = 0.6618843657314521
q40 = 1.830612773966382 q41 = -8.288032061663463
q50 = -0.6245601110692797 q51 = 2.709330688538044
q60 = 0.3850559271659521 q61 = -1.714360699440487
q70 = -0.05705219972778544 q71 = 0.2742239481189889
q02 = -0.4639662012147939 q03 = 0.7228532812132124
q12 = 0.4048092994116597 q13 = - 0.7805039834619138
q22 = 0 q23 = 11.03207597325678
q32 = -1.573371461527950 q33 = 0
q42 = 14.02365705116605 q43 = -10.10065328335495
q52 = -4.343572255913729 q53 = 3.002892932574174
q62 = 2.860630607831792 q63 = - 2.082622244166414
q72 = -0.4973385909142549 q73 = 0.3800614179644372
q04 = - 2.553891864910215 q05 = 2.70983565186940
q14 = 2.234132119994110 q15 = - 2.27133994516961
q24 = -22.47769049395616 q25 = 21.6521061372589
q34 = 2.308946619504642 q35 = - 2.13484805031142
q44 = 0 q45 = 3.77017058952430
q54 = - 1.212267071714534 q55 = 0
q64 = 0.6108231151820283 q65 = - 0.690362730945967
q74 = - 0.06793041339961766 q75 = 0.0990714324530388
q06 = - 1.20540452249925 q07 = 0.1953080412300543
q16 = 1.03696146986553 q17 = -0.1813863173941616
q26 = -10.2885737022032 q27 = 1.956069485925045
q36 = 1.06826234200632 q37 = -0.2131866118947915
q46 = - 1.37062437516525 q47 = 0.1666886279744916
q56 = 0.498101251671513 q57 = -0.07816772353808257
q66 = 0 q67 = 0.8513566887329923
q76 = - 0.778525470347940 q77 = 0
q08 = 0 q09 = 0
q18 = 0 q19 = 0
q28 = 0 q29 = 0
q38 = 0.0007555702610770257 q39 = 0
q48 = -0.03512461897276456 q49 = 0.003305296525213877
q58 = 0.05847352857794642 q59 = -0.01129402979947775
q68 = -0.2873839347920289 q69 = 0.08104365275464091
q78 = 0.8491456051312209 q79 = -0.2627990311996895
 15

q010 = 0 q011 = 0
q110 = 0 q111 = 0
q210 = 0 q211 = 0
q310 = 0 q311 = 0
q410 = 0 q411 = 0
q510 = 0.001062790673425422 q511 = 0
q610 = -0.01565339823898801 q611 = 0.00147301591647824960
q710 = 0.0741105916175047861 q711 = -0.0143142931603533612
q012 = 0
q112 = 0
q212 = 0
q312 = 0
q412 = 0
q512 = 0
q612 = 0
q712 = 0.001347003464450247
16

REFERENCES
1. M.H. Carpenter, D. Gottlieb, and S. Abarbanel. Time-Stable Boundary Conditions
for Finite-Difference Schemes Solving Hyperbolic Systems: Methodology and Appli-
cation to High-Order Compact Schemes. J. Comp. Phys., 111:220–236, (1994).
2. B. Gustafsson, H.-O. Kreiss, and J. Oliger. Time Dependent Problems and Difference
Methods. John Wiley & Sons, New York, (1995).
3. H.-O. Kreiss and G. Scherer. Finite Element and Finite Difference Methods for
Hyperbolic Partial Differential Equations. In C. De Boor, editor, Mathematical As-
pects of Finite Elements in Partial Differential Equations., pages 195–211. Academic
Press, Inc., (1974).
4. B. Strand. Summation by Parts for Finite Difference Approximations for d/dx. J.
Comp. Phys., 110:47–67, (1994).
5. C.K.W Tam and Webb J. C. Dispersion-Relation-Preserving Finite Difference
Schemes for Computational Acoustics. J. Comp. Phys., 107:262–281, (1993).
Paper C
 HIGH ORDER SUMMATION BY PARTS
OPERATOR BASED ON A DRP SCHEME
APPLIED TO 2D AEROACOUSTICS
Stefan Johansson

Division of Scientific Computing

Department of Information Technology
Uppsala University
P.O. Box 337
SE 75105 Uppsala
Sweden

Abstract
A strictly stable high order finite difference method based on Tam
and Webb’s dispersion relation preserving scheme in the interior has been
verified for a 2D aeroacoustic problem. Results show that the method
gives lower dispersion error than a similar method derived by Strand [11],
which is based on standard sixth order difference approximation in the
interior, when boundary effects are not important.

1 Introduction
Numerical simulation of wave propagation problems has been studied us-
ing computers for the last 50 years and is still actively studied. For the
last 10 years computational aeroacoustics (CAA) has emerged as a sub-
field. In CAA sound generation and sound propagation are modeled by
the Navier-Stokes or Euler equations. High accuracy is required due to
long time integration, leading to the need of highly accurate numerical
methods.
Tam and Webb introduced Dispersion Preserving Relation (DRP) sche-
mes [13] that give a better approximation of the wave number than stan-
dard finite difference schemes, thus giving a smaller dispersion error.
One of the more difficult problems for finite difference approximation
is how to apply numerical boundary conditions in a stable way. One ap-
proach proposed by Kreiss and Scherer called summation by parts (SBP)
[7] is to devise methods that give the same energy estimate as for the
continuous problem. Thus, the SBP methods are strictly stable by con-
struction and one does not have to use the algebraically difficult GKS
theory [5] to prove stability.
High accuracy and strict stability have been combined in [6], where
theoretical results and experimental verification were presented for a model
problem in 1D. The main contribution of the present paper is the appli-
cation of the resulting method to aeroacoustic wave propagation in 2D
[12]. The propagation of acoustic, entropy and vorticity waves is com-
puted using the linearized 2D Euler equations. In addition we discuss the
application of numerical dissipation where the damping does not affect
the better resolution of the DRP method compared to a standard finite
difference method.
The outline of the paper is as follows. In Section 2, we introduce the
2D linearized Euler equations to model sound propagation. In Section 3,
the theory for a scalar test case in 1D is recapitulated, using the energy
method to give an energy estimate for the continuous problem and ex-
plaining summation by parts which gives the same estimate in a discrete
norm, cf. [4]. A discussion on how to apply artificial dissipation to Disper-
sion Relation Preserving schemes is given as well. In Section 4 numerical
results that verify the numerical method are presented for a problem in
computational aeroacoustics.

2 2D Linearized Euler Equations

The linearized Euler equations in conservative form [1] are used as a model
of sound propagation. Denoting density ρ, velocity in x-direction u, in y-
direction v, specific total energy E, specific total entalphy H and pressure
p. The equations are formulated in the variables ρ0 , (ρu)0 , (ρv)0 and (ρE)0
that are perturbations of a reference state ρ0 , ρ0 u0 , ρ0 v0 , ρ0 E0 .

   
ρ0 (ρu)0
∂  0 
 (ρu)0  + ∂
 ρ0 u0 u0 + (ρu)0 u0 + p0 
 +
∂t  (ρv)  ∂x  ρ0 v0 u0 + (ρv)0 u0 
(ρE)0 0
ρ0 H0 u + (ρH) u0 0
 
(ρv)0
∂   ρ0 u0 v 0 + (ρu)0 v0 
=0 (1)
∂y  ρ0 v0 v 0 + (ρv)0 v0 + p0 
ρ0 H0 v 0 + (ρH)0 v0

The acoustic quantities are defined by ρ0 = ρ − ρ0 , (ρu)0 = ρu − ρ0 u0 ,

(ρv)0 = ρv − ρ0 v0 and (ρE)0 = ρE − ρ0 E0 . With (ρH)0 = (ρE)0 + p0 .
The variables u0 , v 0 and p0 can be calculated using

(ρu)0 − ρ0 u0
u0 = (2)
ρ0
µ ¶
0 0 1¡ 0 0 ¢
p = (γ − 1) (ρE) − 2 (ρu) · u0 − ρ u0 · u0 (3)
2

with u = (u, v)T

3 Theory
To make the paper self-contained the theory for a scalar test problem in
1D is recapitulated below.

2
3.1 Energy Method
A sufficient condition for well-posedness for an initial boundary value
problem is to give a bound on the solution in terms of the initial value
and boundary data, an energy estimate.
For example, introduce the L2 scalar product and norm
Z 1
(u, v) = uv dx , ||u||2 = (u, u). (4)
0

We have for the simplest hyperbolic equation, sometimes called the Kreiss
equation

 ut = ux , x ∈ [0, 1], t ≥ 0
u(x, 0) = f (x) , x ∈ [0, 1] (5)

u(1, t) = g(t) t ≥ 0
the following energy estimate

d||u||2
= (u, ut )+(ut , u) = (ux , u)+(u, ux ) = [u2 ]10 = g 2 (t)−u2 (0, t). (6)
dt
That leads to an estimate on the norm of u in terms of the initial and
boundary data [4].

3.2 Summation by parts

In the semi-discrete case we want to find an operator Q approximating
d/dx and a discrete norm, kuk2 = (u, u)H = uT Hu, such that the semi-
discretization of (5) gives the same estimate as the continuous case, where
u and v are grid functions on [0, 1] with step size h and H is a diagonal
positive definite matrix. The energy method on ut = Qu gives

d||u||2H
= (u, ut )H +(ut , u)H = (u, Qu)H +(Qu, u)H = uT (HQ+(HQ)T )u
dt
(7)
The same estimate as in the continuous is given if (u, Qu)H + (Qu, u)H =
[u2 ]10 , or equivalent HQ + (HQ)T = diag(−1, 0, . . . , 0, 1). This property
is called summation by parts (SBP).
Operators with this property were proposed by Kreiss and Scherer
[7] and extended to high order by Strand [11]. In [6] SBP operators
for difference schemes of Dispersion Relation Preserving (DRP) schemes
described in Section 3.3 type were developed.

3.3 Dispersion Relation Preserving Schemes

If the formal accuracy of a finite difference method is lowered by two
orders as proposed by Tam and Webb [13] a free parameter is given that
can be chosen such that the wave number is approximated better.
That free parameter φ is chosen such that error E(φ) in the approxi-
mation of the wave number is minimized
Z π/2
E(φ) = |hk − hk̃(φ)|2 d(hk) (8)
−π/2

where k and k̃ are the exact and the approximate wave numbers, respec-
tively.

3
 A comparison between standard centered finite difference methods and
the DRP schemes is given in Figure (1). The straight line shows the exact
non-dimensional wave number hk. The dotted, dotted-dashed and dashed
lines show the approximate non-dimensional wave numbers hk̃ for different
schemes. As seen the fourth order DRP scheme derived from the sixth
order standard centered difference method gives a good approximation for
the non-dimensional wave numbers up to about π/2 whereas the standard
sixth order method gives a good approximation up to about 1.2. The
details of the present fourth order DRP scheme and 6th, 8th order DRP
schemes can be found in [6].

3.5

2.5
Approximate wave number

1.5

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5
kh

Figure 1: Approximate non-dimensional wave number vs exact for 2nd order

standard centered difference method (SC2) (dotted), SC6 (dot-dashed), 4th
order DRP scheme (dashed).

3.4 Numerical Dissipation

To damp high frequency modes in the numerical solution that are under-
resolved artificial dissipation is often applied. The standard explicit type
is (−1)p+1 β(h2 D+ D− )p , where β > 0 and p ≥ 1.
But since DRP schemes are derived so that they give a good approxi-
mation for wave numbers less that π/2, very high order dissipation has to
be used not to destroy that property. Figure 2 shows the Fourier trans-
form of low pass filters for p = 1 . . . 4 which demonstrate that the damping
is significant for wave numbers < π/2. Alternatives are to optimize the
stencil in the artificial dissipation analogously to to use the idea proposed
for filters by [2] or using an implicit artificial dissipation similar to the

4
 idea used for filters proposed by [14].

1+p/2 p
Damping charcteristics of V + (−1/2) δ V
1
p=2
p=4
0.9 p=6
p=8

0.8

0.7

0.6
Fourier transform

0.5

0.4

0.3

0.2

0.1

0
0 0.5 1 1.5 2 2.5 3 3.5
k∆x

Figure 2: Fourier transform of low pass filters of orders 2, 4, 6, 8.

Assume that S is a dissipation operator, the energy method on the

Kreiss equation in semi discrete form with dissipation added ut = Qu +
H −1 Su gives

d
kukH = (ut , u)H + (u, ut )H = (Qu + Su, u)H + (u, Qu + Su)H =
dt
uT (HQ + (HQ)T + S + S T )u (9)

If the interior stencil of S is (−1)p+1 β(h2 D+ D− )p a boundary closure that

guarantees that S + S T leq0 can be found [3] [8].
Using an implicit dissipation (−1)p+1 βP −1 (h2 D+ D− )p , where P is a
symmetric tridiagonal matrix with ones on the diagonal does not give
an energy estimate with the standard boundary closure for p = 1, 2, 3,
which was verified numerically. In general nothing can be said about the
definiteness of P −1 S.

4 Application to a 2D aeroacoustic prob-

lem
The high order SBP operator based on the Tam and Webb DRP scheme
was implemented in a FORTRAN program [10] and tested for an aeroa-

5
coustic test problem. The results were compared with those obtained with
a SBP operator based on a standard central finite difference stencil.
The 2D linearized Euler equations described in Section 2 were solved
using those SBP operators to discretize x and y derivatives and using
the standard fourth order Runge-Kutta method to discretize the time
derivative.
A benchmark problem was chosen from [12], choosing u0 = 0.5, v0 = 0
with the following initial condition

x2 +y 2
p0 (x, y, 0) = e−log(2) 9 (10)
2 2 (x−67)2 +y 2
−log(2) x +y
ρ0 (x, y, 0) = e 9 + 0.1e−log(2) 25 (11)
(x−67)2 +y 2
u0 (x, y, 0) = 0.04ye −log(2) 25 (12)
(x−67)2 +y 2
0 −log(2)
v (x, y, 0) = −0.04(x − 67)e 25 (13)

to the linearized Euler equations. There exists an analytical solution

for verification of the numerical results [12].
Numerical solution obtained with the DRPSBP scheme for ρ0 , the
acoustic density, with the two resolutions at time instant T = 40 can be
found in Figure 3, near the center the acoustic pulse is seen, propagat-
ing with the speed of sound in all directions transported with the mean
velocity. Near the right boundary the entropy and vorticity wave is seen
propagating with the mean velocity.
The time instances were deliberately chosen such that the solution had
not reached the boundary, since the non reflecting boundary condition
implemented is not of high order and the purpose of this paper is to
evaluate the new SBP method. Results for the first order Enquist-Majda
boundary condition can be found in [10]
Figure 4 to Figure 7 show the numerical solution using SBP-3-6 (SBP
operator based on the sixth order standard finite difference scheme with
third order near the boundaries) compared to DRPSBP-2-4 (SBP operator
based on a fourth order DRP scheme derived from a sixth order standard
finite difference scheme with second order near the boundaries) of ρ0 along
the x-axis at time T = 40, compared to the exact solution for two different
grid sizes, and with and without filtering. The filter is applied in each time
step and of standard type (−1)p+1 2−2p (h2 D+ D− )p , in this case p = 3 is
chosen [9], with the boundary treatment according to [3].
Figure 4 shows that the method based on a DRP scheme has lower dis-
persion error and has more accurate peak values, and smaller amplitudes
on the wiggles but more noticeable overshoots. The entropy and vorticity
wave at x = 87 is better resolved than the acoustic waves at x = −20 and
x = 60. Table 1 and Table 5 show that DRPSBP-2-4 has about a 60 %
lower L2 -error along the x-axis for the lower resolution.
When refining the grid the differences between the methods are not
seen in Figure 5. Table 2 and Table 6 show that the difference in L2 -error
along the x-axis is about 10 %, but to advantage of the SBP method.
On the other hand when applying a standard type filter to damp the
high frequency oscillations, there is less difference between the methods
as shown in Figure 6 and Figure 7. Table 3, Table 4, Table 7 and Table
8 stating the L2 -error along the x-axis for the two grid sizes and the
two time instances chosen, demonstrate that the L2 -error along the x-
axis is larger than without filtering. The reason is that the filter imposes

6
 significant damping for wave numbers hk < π/2 cf. Figure 2, which affects
the higher resolution of the DRP based SBP method.

DRPSBP24, T=40, NlK, N=100, ρ

0.1

0.08

0.06

0.04

0.02

−0.02

−0.04

−0.06
100

50 100
50
0
0
−50
−50
−100 −100
y
x

Figure 3: Surface plot of acoustic density at time 40, without filtering, top:
N=101, bottom: N=201.

7
 DRPSBP24, T=40, NlK, N=100, ρ
0.1
Exact
Numerical
0.08

0.06

0.04

0.02
ρ

−0.02

−0.04

−0.06
−100 −80 −60 −40 −20 0 20 40 60 80 100
x

SBP36, T=40, NlK, N=100, ρ

0.1
Exact
Numerical
0.08

0.06

0.04

0.02
ρ

−0.02

−0.04

−0.06

−0.08
−100 −80 −60 −40 −20 0 20 40 60 80 100
x

Figure 4: Numerical solution of ρ0 at time 40, N = 101, without filtering, top:

DRPSBP-2-4, bottom: SBP-3-6.

8
 DRPSBP24, T=40, NlK, N=200, ρ
0.12
Exact
Numerical
0.1

0.08

0.06

0.04
ρ

0.02

−0.02

−0.04

−0.06
−100 −80 −60 −40 −20 0 20 40 60 80 100
x

SBP36, T=40, NlK, N=200, ρ

0.12
Exact
Numerical
0.1

0.08

0.06

0.04
ρ

0.02

−0.02

−0.04

−0.06
−100 −80 −60 −40 −20 0 20 40 60 80 100
x

Figure 5: Numerical solution of ρ0 at time 40, N = 201, without filtering, top:

DRPSBP-2-4, bottom: SBP-3-6.

9
 DRPSBP24, T=40, 0lK, N=100, ρ
0.1
Exact
Numerical
0.08

0.06

0.04

0.02
ρ

−0.02

−0.04

−0.06
−100 −80 −60 −40 −20 0 20 40 60 80 100
x

SBP36, T=40, 0lK, N=100, ρ

0.1
Exact
Numerical
0.08

0.06

0.04

0.02
ρ

−0.02

−0.04

−0.06
−100 −80 −60 −40 −20 0 20 40 60 80 100
x

Figure 6: Numerical solution of ρ0 at time 40, N = 101, sixth order filter, top:
DRPSBP-2-4, bottom: SBP-3-6.

10
 DRPSBP24, T=40, 0lK, N=200, ρ
0.1
Exact
Numerical
0.08

0.06

0.04

0.02
ρ

−0.02

−0.04

−0.06
−100 −80 −60 −40 −20 0 20 40 60 80 100
x

SBP36, T=40, 0lK, N=200, ρ

0.1
Exact
Numerical
0.08

0.06

0.04

0.02
ρ

−0.02

−0.04

−0.06
−100 −80 −60 −40 −20 0 20 40 60 80 100
x

Figure 7: Numerical solution of ρ0 at time 40, N = 201, sixth order filter, top:
DRPSBP-2-4, bottom: SBP-3-6.

11
 DRPSBP-2-4 SBP-3-6
L2-error along the x-axis in ρ 0.00431254 0.00788948
L2-error along the x-axis in p 0.00431147 0.00788902
L2-error along the x-axis in u 0.00428112 0.00649726

Table 1: u0 =0.5, T=20, N = 101, no filter

DRPSBP-2-4 SBP-3-6
L2-error along the x-axis in ρ 0.00049776 0.00047037
L2-error along the x-axis in p 0.00049762 0.00047037
L2-error along the x-axis in u 0.00049358 0.00046720

Table 2: u0 =0.5, T=20, N = 201, no filter

DRPSBP-2-4 SBP-3-6
L2-error along the x-axis in ρ 0.01022241 0.01020396
L2-error along the x-axis in p 0.01018905 0.01017082
L2-error along the x-axis in u 0.00991698 0.00986553

Table 3: u0 =0.5, T=20, N = 101, sixth order filter

DRPSBP-2-4 SBP-3-6
L2-error along the x-axis in ρ 0.00118599 0.00114419
L2-error along the x-axis in p 0.00118562 0.00114392
L2-error along the x-axis in u 0.00117612 0.00113471

Table 4: u0 =0.5, T=20, N = 201, sixth order filter

DRPSBP-2-4 SBP-3-6
L2-error along the x-axis in ρ 0.00444214 0.00723430
L2-error along the x-axis in p 0.00436382 0.00723236
L2-error along the x-axis in u 0.00372037 0.00676451

Table 5: M=0.5, T=40, N = 101, no filter

DRPSBP-2-4 SBP-3-6
L2-error along the x-axis in ρ 0.00074894 0.00063326
L2-error along the x-axis in p 0.00071616 0.00063313
L2-error along the x-axis in u 0.00071466 0.00063204

Table 6: u0 =0.5, T=40, N = 201, no filter

12
 DRPSBP-2-4 SBP-3-6
L2-error along the x-axis in ρ 0.00903172 0.00910053
L2-error along the x-axis in p 0.00885448 0.00895898
L2-error along the x-axis in u 0.00879172 0.00881355

Table 7: u0 =0.5, T=40, N = 101, sixth order filter

DRPSBP-2-4 SBP-3-6
L2-error along the x-axis in ρ 0.00148481 0.00138574
L2-error along the x-axis in p 0.00147850 0.00138488
L2-error along the x-axis in u 0.00147532 0.00138194

Table 8: u0 =0.5, T=40, N = 201, sixth order filter

The numerical results for the 2D linearized Euler equations show that
the better approximation of the wave number for the DRP based SBP
operator of fourth order (second order near the boundary) results in a
smaller L2 -error than for a SBP operator based on standard sixth order
(third order near the boundary) stencil. The results are valid when the
problem is not fully resolved, which usually is the case for large scale
applications.
Application of a sixth order filter affects the good wave approximation
of the DRPSPB operator. That problem can be solved by using a filter or
artificial dissipation that apply less dissipation for wave numbers hk < π/2
than the present filter.

5 Conclusions and Future Work

A strictly stable finite difference method with the summation by parts
property (SBP) based on Tam and Webb’s dispersion relation preserving
(DRP) scheme in the interior developed in [6] has been tested on a 2D
aeroacoustic problem. The results show that the SBP method based on the
DRP scheme yields smaller errors for low resolutions without using a filter.
But the advantage disappears with high resolution and/or using a sixth
order explicit filter. However, using other types of artificial dissipation or
filters, that are tailored to match the better wave number resolution of the
DRP scheme, should lead to better results. Also in realistic applications
it is often not possible to use the resolution needed, which shows the
advantage of the DRPSBP method. The development and implementation
of accurate non reflecting boundary conditions is future work. We also
plan to develop high order numerical dissipation that will replace the filter
used in this paper.

13
References
[1] M. Billson, L.-E. Eriksson, and L. Davidson. Acoustic Source
Terms for the Linear Euler Equations on Conservative Form. In
8th AIAA/CAES Aeroacoustics Conference, Breckenridge, Colorado,
(2002). AIAA Paper 2002-2582.
[2] C. Bogey and C. Bailly. A family of low dispersive and low dissipative
explicit schemes for flow and noise computations. J. Comp. Phys.,
194(1):194–214, (2004).
[3] L.-E. Eriksson. Boundary conditions for artificial dissipation opera-
tors. Technical Report FFA TN 1984-53, The Aeronautical Reaserch
Institute of Sweden, (1984).
[4] B. Gustafsson, H.-O. Kreiss, and J. Oliger. Time Dependent Problems
and Difference Methods. John Wiley & Sons, New York, (1995).
[5] B. Gustafsson, H.-O. Kreiss, and A. Sundström. Stability Theory of
Difference Approximations for Mixed Initial Boundary Value Prob-
lems. II. Math. Comp., 26(119):649–686, (1972).
[6] S. Johansson. High Order Finite Difference Operators with the Sum-
mation by Parts Property Based on DRP Schemes. Technical Report
2004-035, Uppsala University, Department of Scientific Computing,
2004.
[7] H.-O. Kreiss and G. Scherer. Finite Element and Finite Differ-
ence Methods for Hyperbolic Partial Differential Equations. In
C. De Boor, editor, Mathematical Aspects of Finite Elements in Par-
tial Differential Equations., pages 195–211. Academic Press, Inc.,
(1974).
[8] K. Mattsson, M. Svärd, and J. Nordström. Stable and Accurate
Artificial Dissipation. J. Sci. Comp., 21(1):57–79, (2004).
[9] B. Müller and S. Johansson. Strictly Stable High Order Difference
Approximations for Low Mach Number Aeroacoustics. In P. Neit-
taanmäki, T. Rossi, K. Majava, and O. Pironneau, editors, ECCO-
MAS 2004, 2004.
[10] C. L. Müller. High order accurate numerical solution of the lin-
earized Euler equations for sound propagation in the atmosphere.
Master’s thesis, Uppsala University, Department of Scientific Com-
puting, (2004).
[11] B. Strand. Summation by Parts for Finite Difference Approximations
for d/dx. J. Comp. Phys., 110:47–67, (1994).
[12] C.K.W. Tam. Benchmark Problems and Solutions. In J.C. Hardin,
J.R. Ristorcelli, and C.K.W. Tam, editors, ICASE/LaRCWorkshop
on Benchmark Problems in Computational Aeroacoustics (CAA),
(1995). NASA CP 3300.
[13] C.K.W Tam and J. C. Webb. Dispersion-Relation-Preserving Finite
Difference Schemes for Computational Acoustics. J. Comp. Phys.,
107:262–281, (1993).
[14] M. R. Visbal and D. V. Gaitonde. High-Order-Accurate Methods
for Complex Unsteady Subsonic Flows. AIAA J., 37(10):1231–1239,
(1999).

14