Geophys. J. Int. (2004) 157, 1269–1296 doi: 10.1111/j.1365-246X.2004.02289.

Mixed-grid and staggered-grid finite-difference methods

for frequency-domain acoustic wave modelling

Bernhard Hustedt, Stéphane Operto and Jean Virieux

UMR Géosciences Azur, CNRS-UNSA-UPMC, France. E-mails: [email protected]; [email protected]; [email protected]

Accepted 2004 March 12. Received 2004 February 10; in original form 2002 June 21

SUMMARY
We compare different finite-difference schemes for two-dimensional (2-D) acoustic frequency-
domain forward modelling. The schemes are based on staggered-grid stencils of various accu-
racy and grid rotation strategies to discretize the derivatives of the wave equation. A combina-
tion of two O(x 2 ) staggered-grid stencils on the classical Cartesian coordinate system and
the 45◦ rotated grid is the basis of the so-called mixed-grid stencil. This method is compared
with a parsimonious staggered-grid method based on a fourth-order approximation of the first
derivative operator. Averaging of the mass acceleration can be incorporated in the two stencils.
Sponge-like perfectly matched layer absorbing boundary conditions are also examined for each
stencil and shown to be effective.
The deduced numerical stencils are examined for both the wavelength content and az-
imuthal variation. The accuracy of the fourth-order staggered-grid stencil is slightly superior
in terms of phase velocity dispersion to that of the mixed-grid stencil when averaging of the
mass acceleration term is applied to the staggered-grid stencil.
For fourth-order derivative approximations, the classical staggered-grid geometry leads
to a stencil that incorporates 13 grid nodes. The mixed-grid approach combines only nine

GJI Seismology
grid nodes. In both cases, wavefield solutions are computed using a direct matrix solver based
on an optimized multifrontal method. For this 2-D geometry, the staggered-grid strategy is
significantly less efficient in terms of memory and CPU time requirements because of the
enlarged bandwidth of the impedance matrix and increased number of coefficients in the
discrete stencil.
Therefore, the mixed-grid approach should be suggested as the routine scheme for 2-D
acoustic wave propagation modelling in the frequency domain.
Key words: derivative stencils, finite-difference methods, frequency-domain wave mod-
elling, seismic wave propagation.

1 I N T RO D U C T I O N

Modelling seismic wave propagation is essential for understanding complex wave phenomena in a realistic heterogeneous medium. Numerical
results from finite-difference (FD) modelling are particularly useful since they provide the complete wavefield response. Frequency-domain
forward modelling is of special interest for multisource experiments, such as tomographic experiments, because of its computational efficiency
(Pratt & Worthington 1990; Štekl & Pratt 1998). Moreover, realistic rheology is easily incorporated into the modelling scheme by introducing
complex velocities. The key step in frequency-domain finite-difference (FDFD) modelling that controls computational efficiency is the
numerical inversion of a massive matrix equation. The matrix structure depends on the spatial derivative approximations. We shall discuss
what the useful features of this matrix are for wave modelling accuracy and for computational efficiency.
Elastodynamic finite-difference time-domain (FDTD) techniques moved from second-order approximations of spatial derivatives
(Madariaga 1976; Virieux 1984, 1986) to higher-order approximations (Dablain 1986; Levander 1988) using staggered-grid stencils with
a good trade-off between modelling accuracy and computational efficiency for the fourth-order spatial approach. Constructions of FDFD
schemes have followed a similar path. Pratt & Worthington (1990) and Pratt (1990) have applied second-order spatial derivative approximations
in two dimensions to the acoustic and elastic wave equations using centred finite-difference (FD) discretizations on a single grid. This leads
to five-point and nine-point stencils for the acoustic and elastic wave equations respectively. Their scheme required 10 gridpoints per shortest
wavelength for accurate modelling. Jo et al. (1996) reduced the number of gridpoints per shortest wavelength to about four by including
additional surrounding gridpoints for the derivative approximations in the acoustic wave equation. The approach consists of linearly combining


C 2004 RAS 1269
1270 B. Hustedt, S. Operto and J. Virieux

two discretizations of the second derivative operator on the classical Cartesian coordinate system and the 45◦ rotated system. Combination
of the two discretizations mitigates the anisotropy of the stencil but increases its numerical dispersion. Additional weighted averaging of the
mass acceleration term over the gridpoints of the FD stencil gives a compact stencil with both optimal anisotropy and dispersion properties.
The weight coefficients which are required to mix the two different derivative stencils and average the acceleration term are derived by an
optimization technique that minimizes the phase velocity dispersion. The approach of Jo et al. (1996) will be referred to in the following as
the mixed-grid method.
The mixed-grid approach was extended to the elastic wave equation by Štekl & Pratt (1998). In the elastic case, the approach is more
complicated since vectorial fields must be manipulated. Moreover, the weight coefficients depend on the Poisson ratio. In the fluid case, the
mixed-grid stencil must be adapted such that only the discretization on the rotated frame is used (Štekl & Pratt 1998).
A second extension of the mixed-grid method was developed by Shin & Sohn (1998). They applied a 25-point stencil to the acoustic
wave equation. Four directions (0◦ , 26.6◦ , 45◦ and 63.4◦ ) were considered to discretize the Laplacian. The number of gridpoints per shortest
wavelength was reduced to two and a half instead of four for the nine-point mixed-grid stencil. For these discretization rules, the bandwidth
of the impedance matrix remains the same but the number of core memory elements is reduced by a factor of four when using the 25-point
stencil (Shin & Sohn 1998). One drawback of the 25-point stencil may be that structures whose size is smaller than one wavelength and
complex topographies cannot be properly discretized when using two and a half points per wavelength. Moreover, the source implementation
on a coarse grid may lead to some inaccuracies.
While time-domain approaches rely strongly on the staggered-grid structure, the frequency-domain formulations were developed on
non-staggered-grid strategies. We shall analyse why different strategies have been developed. In order to do so, we focus our attention on the
acoustic wave equation.
In the following, we first present a parsimonious second-order staggered-grid formulation of the acoustic wave equation in the frequency
domain following Luo & Schuster (1990). The acoustic wave equation is formulated as a first-order hyperbolic system (Virieux 1984) involving
fields as the pressure and the particle velocities. This system is discretized using second-order staggered-grid stencils. Then, particle velocities
are eliminated from the coupled first-order equations. Elimination of particle velocities leads to the so-called parsimonious second-order
staggered-grid formulation. This parsimonious formulation only involves the pressure field in the second-order wave equation.
We shall show the identity between the parsimonious second-order staggered-grid stencil and the five-point stencil originally developed
by Pratt & Worthington (1990). We also develop the second-order staggered-grid stencil on the 45◦ rotated coordinate system following the
same parsimonious elimination procedure. This rotated stencil construction was recently proposed by Saenger et al. (2000) for time-domain
modelling of the elastic wave equation.
The two second-order derivations provide a general framework within which to discretize the frequency-domain second-order acoustic
wave equation based on a parsimonious staggered-grid approach. A sponge-like perfectly matched layer (PML) absorbing boundary condition
(Berenger 1994) is also developed for both stencils.
In Section 3 we present two improved discretizations of the frequency-domain wave equation. The first relies on the approach of Jo
et al. (1996). We clarify the close connection between the staggered-grid and the mixed-grid method: the mixed-grid method combines two
second-order staggered-grid stencils to mimic a fourth-order accurate FD stencil. We also provide a qualitative interpretation to justify why
the acceleration averaging in the mixed-grid method is a key component of the method. The second discretization extends to fourth-order
accuracy the parsimonious second-order staggered-grid method. This discretization leads to a 13-point stencil.
In Section 4 the dispersion properties of the mixed-grid and fourth-order staggered-grid stencils are compared through a classical
harmonic analysis. We show that the accuracy of the fourth-order staggered-grid stencil is slightly better than the mixed-grid stencil when
averaging of the mass acceleration term is applied to the staggered-grid stencil.
In Section 5 we compare the structure of the impedance matrices of the two stencils. The CPU and RAM performances of the mixed-grid
method are expected to be significantly superior to those of the fourth-order staggered-grid method. This results from the enlargement of
the matrix bandwidth when using the 13-point stencil associated with the fourth-order staggered-grid method. Due to their compactness, the
mixed-grid stencils appear to be optimal in terms of CPU and RAM performances, as has already been demonstrated by Štekl & Pratt (1998).
In Section 6 we present two numerical examples to quantify the accuracy and CPU performances of the mixed-grid stencil and the
fourth-order staggered-grid stencil.

2 T H E S E C O N D - O R D E R F I N I T E - D I F F E R E N C E D I S C R E T I Z AT I O N

2.1 The first staggered-grid strategy

In seismology the staggered-grid strategy has so far been developed mainly in the time domain (Madariaga 1976; Virieux 1984, 1986). The
staggered leap-frog structure of the numerical grid guarantees accurate derivative estimations and good performance for wave propagation
dispersion.
We formulate the acoustic wave equation as a first-order hyperbolic system (Virieux 1984) using physical quantities such as the pressure
and the particle velocity. Sponge-like PML at the edges of the necessary limited numerical model are implemented to absorb energy (Berenger


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1271

1994). The system of hyperbolic equations is given by:

∂ Px (x, z, t) ∂ Q(x, z, t)
+ γx (x)Px (x, z, t) = K (x, z) + S(x, z, t)
∂t ∂x
∂ Pz (x, z, t) ∂ R(x, z, t)
+ γz (z)Pz (x, z, t) = K (x, z)
∂t  ∂z 
∂ Q(x, z, t) ∂ Px (x, z, t) ∂ Pz (x, z, t)
+ γx (x)Q(x, z, t) = b(x, z) +
∂t ∂x ∂x
 
∂ R(x, z, t) ∂ Px (x, z, t) ∂ Pz (x, z, t)
+ γz (z)R(x, z, t) = b(x, z) + (1)
∂t ∂z ∂z
where wavefields Q(x, z, t) and R(x, z, t) are components of the particle velocity vector on the Cartesian grid while the pressure P(x, z, t) is split
into two components, P x (x, z, t) and P z (x, z, t), for separation of horizontal and vertical derivatives. The pressure wavefield is deduced through
the simple addition of the two unphysical acoustic fields P x and P z , P(x, z, t) = P x (x, z, t) + P z (x, z, t). The two unphysical acoustic fields
P x and P z are used to account for the PML absorbing boundary conditions (see Zhang & Ballmann 1997; Operto et al. 2002, for application
of PML boundary conditions to the SH and P-SV wave equations). S(x, z, t) is the pressure source term applied to the sum of the first two
equations of eqns (1) outside the PML layers (in that case, the γ functions are zero):
∂ P(x, z, t) ∂ Q(x, z, t) ∂ R(x, z, t)
= K (x, z) + K (x, z) + S(x, z, t). (2)
∂t ∂x ∂z
Since the pressure source term is an additive term, it can be equivalently applied either to the first or second equation of system (1).
The inverse of the density is denoted by the buoyancy b(x, z). If the bulk modulus K(x, z) grid defines the reference grid (i.e. the
grid input in the numerical code), the density (or the buoyancy) needs to be interpolated between gridpoints horizontally and vertically (see
the classical geometry in Fig. 1). We used harmonic averaging of the density to compute such an interpolation (Kelly et al. 1976; Graves
1996).
The 1-D functions γ x and γ z define the PML damping behaviour in the PML layers surrounding the model on all four sides. These
functions differ from zero only inside the PML layers. In the PML layers, we used γ (x) = cPML cos( π2 Lx ) where L denotes the width of the
PML layer and x is a local coordinate in the PML layer whose origin is located at the outer edges of the model. The scalar c PML is defined by
trial and error depending on the width of the PML layer (i.e. several values of c PML are tested for representative PML widths; the best value
is the one for which the reflections coming from the edges of the model have the smallest amplitude).

Two staggered-grid geometries

classic rotated
{
, x’

R1 R2

R3 R4
z’

Figure 1. Illustration of different staggered finite-difference grids. Besides the pressure field, staggered-grid formulations of the acoustic wave equation
introduce two additional variables, Q and R. The first classical discretization strategy gives different positions for these variables while the second alternative
rotation strategy requires these variables to be located at the same grid node position.


C 2004 RAS, GJI, 157, 1269–1296
1272 B. Hustedt, S. Operto and J. Virieux

We transform the system of equations to the Fourier domain and introduce the new functions ξ x (x) = 1 + iγ x (x)/ω and ξ z (z) = 1 +
iγ z (z)/ω to simplify the equations:

−iωξx (x) ∂ Q(x, z, ω)

Px (x, z, ω) = + S(x, z, ω)
K (x, z) ∂x
−iωξz (z) ∂ R(x, z, ω)
Pz (x, z, ω) =
K (x, z) ∂z
b(x, z) ∂ P(x, z, ω)
−iωQ(x, z, ω) =
ξx (x) ∂x
b(x, z) ∂ P(x, z, ω) (3)
−iω R(x, z, ω) = .
ξz (z) ∂z

We discretize the hyperbolic equation system (3) using second-order centred finite differences of partial derivatives:
 
∂Q 1
≈ (Q i+1/2, j − Q i−1/2, j )
∂ x i, j 
 
∂R 1
≈ (Ri, j+1/2 − Ri, j−1/2 ), (4)
∂z i, j 

where  is the mesh spacing in the uniform grid. We obtain the following discrete system of equations:

−iωξxi 1
Pxi, j = (Q i+1/2, j − Q i−1/2, j ) + Si, j
K i, j 
−iωξz j 1
Pzi, j = (Ri, j+1/2 − Ri, j−1/2 )
K i, j 
bi+1/2, j 1
−iωQ i+1/2 = (Pi+1, j − Pi, j )
ξxi+1/2 
bi, j+1/2 1
−iω R j+1/2 = (Pi, j+1 − Pi, j )
ξz j+1/2 
bi−1/2, j 1
−iωQ i−1/2 = (Pi, j − Pi−1, j )
ξxi−1/2 
bi, j−1/2 1
−iω R j−1/2 = (Pi, j − Pi, j−1 ), (5)
ξz j−1/2 

where we have introduced the averaged coefficients bi+1/2, j = 12 (bi+1, j + bi, j ), and ξxi+1/2 = 12 (ξxi+1 + ξxi ).
We eliminate the additional equations depending on Q and R by re-injecting them into the first two equations that depend on P x and P z
only, and obtain the compact system of equations:
 
−iωξxi 1 1 bi+1/2, j 1 1 bi−1/2, j 1
Pxi, j = − (Pi+1, j − Pi, j ) + (Pi, j − Pi−1, j ) + Si, j
K i, j  iω ξxi+1/2  iω ξxi−1/2 
 
−iωξzi 1 1 bi, j+1/2 1 1 bi, j−1/2 1
Pzi, j = − (Pi, j+1 − Pi, j ) + (Pi, j − Pi, j−1 ) . (6)
K i, j  iω ξz j+1/2  iω ξzi−1/2 

The procedure for eliminating the Q and R fields in the discretized equations of system (6) defines the so-called parsimonious approach (Luo
& Schuster 1990).
We sum the two unphysical wavefields, P x and P z , and get the second-order finite-difference equation with PML absorbing boundary
conditions for the pressure wavefield P:
 
−ω2 1 1 bi+1/2, j bi−1/2, j
Pi, j = (Pi+1, j − Pi, j ) − (Pi, j − Pi−1, j )
K i, j ξxi 2 ξxi+1/2 ξxi−1/2
 
1 1 bi, j+1/2 bi, j−1/2
+ (Pi, j+1 − P i, j ) − (P i, j − Pi, j−1 ) + Si, j . (7)
ξz j 2 ξz j+1/2 ξz j−1/2

Note that although only one scalar field P is now considered after elimination, the underlying staggered-grid structure still appears in the
parsimonious formulation, eq. (7), through the estimation of the buoyancy parameter at intermediate positions with respect to the pressure
gridpoints. Buoyancy at intermediate positions labelled by indices i ± 1/2, j ± 1/2 is computed by arithmetic averaging.


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1273

We extract the coefficients that correspond to the staggered-grid five-point stencil (see Fig. 1):
   
ω2 1 bi+1/2, j bi−1/2, j 1 bi, j+1/2 bi, j−1/2
C1 = − + − +
K i, j ξxi 2 ξxi+1/2 ξxi−1/2 ξz j 2 ξz j+1/2 ξz j−1/2
1 bi−1/2, j
C2 =
ξxi 2 ξxi−1/2
1 bi+1/2, j
C3 =
ξxi 2 ξxi+1/2
1 bi, j−1/2
C4 =
ξz j 2 ξz j−1/2
1 bi, j+1/2 (8)
C5 = .
ξz j 2 ξz j+1/2

We have provided the procedure to derive parsimonious staggered-grid FD stencils for the frequency-domain second-order wave equation.
In order to analyse the properties of the staggered-grid stencils from a qualitative viewpoint, we now remind the reader of an intrinsic feature
of staggered-grid stencils.
Let us assume we want to perform FDTD wave propagation modelling using the time-domain first-order hyperbolic formulation, equation
system (1). For the purposes of illustration, the system is discretized for variables P, Q and R on a single non-staggered grid using a centred
finite-different stencil (−1/2, 0, 1/2) for the first-derivative operator. By a non-staggered grid, we define here the single grid built by adding
all intermediate points to the classical staggered grids of Fig. 1. If we denote by  ns the mesh spacing of the non-staggered grid, then
the mesh spacing of the staggered grids would be  = 2 ns . The coefficient 0 in the centred FD stencil (−1/2, 0, 1/2) accounts for the
non-staggered-grid structure (i.e. by construction, this zero coefficient does not exist when staggered grids are considered). The coefficients
±1/2 account for the spatial support 2 ns of the centred stencil, used to compute the first-order derivative.
We consider a homogeneous medium with a velocity of 4000 m s−1 and a density of 2500 kg m−3 . A point source is located in the middle
of the medium. The source is excited at one single grid node. A time-domain snapshot of the pressure field shows two uncoupled staggered
grids: one that is not excited (i.e. the amplitude of the pressure nodes was left to zero) while the other is related to the source position and will
provide the simulation we are interested in Fig. 2(a). Snapshots of the two velocity fields (not shown there) also exhibit uncoupled staggered
grids, except that the nodes of the two subgrids are not at the same location in the pressure and velocity grids. One may remove the unexcited
subgrids in the pressure and velocity grids leading to the first staggered-grid geometry of Fig. 1. The decoupling between the staggered grids
shows that the staggered-grid stencil is accurate to second-order in the space increment 2 ns . This analysis will be applied later to other
stencils to compare their accuracy qualitatively.

2.2 Comparison with the classical grid approach

A standard five-point stencil was constructed directly by Pratt & Worthington (1990, eqs B6 and B7). They discretized the second-order
hyperbolic acoustic wave equation:
   
∂ 1 ∂ ∂ 1 ∂ ω2
P(x, z, ω) + P(x, z, ω) + P(x, z, ω) = S(x, z, ω), (9)
∂ x ρ(x, z) ∂ x ∂z ρ(x, z) ∂z K (x, z)
using the finite-difference approximation of spatial derivatives proposed by Boore (1972, eq. 11) or by Kelly et al. (1976). These finite-difference
operators are given by:
    
∂ 1 ∂ 1 1 1
P(x, z, ω) ≈ 2 (Pi+1, j − Pi, j ) − (Pi, j − Pi−1, j )
∂ x ρ(x, z) ∂ x i, j  ρi+1/2, j ρi−1/2, j
    
∂ 1 ∂ 1 1 1
P(x, z, ω) ≈ 2 (Pi, j+1 − Pi, j ) − (Pi, j − Pi, j−1 )
∂z ρ(x, z) ∂z i, j  ρi, j+1/2 ρi, j−1/2 (10)
where
 
1 1 1 1
= + .
ρi+1/2, j 2 ρi+1, j ρi, j
An immediate comparison shows that this standard five-point stencil is identical to the previous staggered-grid five-point stencil (eq. 7) away
from the PML zones where damping functions ξ (x) and ξ (z) are equal to 1.
This analysis shows the identity of the second-order numerical schemes between the staggered-grid and the standard-grid formulations
as long as harmonic averaging of density is considered in the standard-grid formulation.

2.3 The alternative rotated staggered-grid strategy

One may consider that the spatial derivatives of the system of eqns (1) could be discretized along rotated axes of a system (x , z )
(Fig. 1). Let us underline that this is not equivalent to a complete rotation because wavefield components are still considered in the


C 2004 RAS, GJI, 157, 1269–1296
1274 B. Hustedt, S. Operto and J. Virieux

(a) Distance (m) Distance (m)

0 1000 2000 3000 4000 900 910 920 930 940 950 960 970 980 990 1000
0 1400

1410

1420
1000
1430

1440
Depth (m)

Depth (m)
2000 1450

1460

1470
3000
1480

1490

4000 1500

(b) 0 1000
Distance (m)
2000 3000 4000
Distance (m)
900 910 920 930 940 950 960 970 980 990 1000
0 1400

1410

1420
1000
1430

1440
Depth (m)

Depth (m)

2000 1450

1460

1470
3000
1480

1490

4000 1500

Figure 2. For a non-staggered-grid formulation we show on the left side a snapshot of the pressure field propagating in a homogeneous medium and excited
by an impulsive point source. On the right side we show a close-up of this snapshot. The intersections between the dashed lines indicate the locations of the
gridpoints. (a) Second-order staggered-grid stencil of coefficients (−1/2, 0, 1/2) for the first derivative is applied: two subgrids are superposed. Only one grid
is excited as shown by the zoom. Grey nodes are not excited and could be√ eliminated.√The white square on the zoom delineates the associated staggered-grid
cell. (b) Second-order rotated staggered-grid stencil of coefficients (−1/2 2, 0, 1/2 2) is applied. Two subgrids are still superposed. The white square still
denotes the elementary staggered-grid cell. Note that contrary to the case shown in (a) two gridpoints (the upper left and the bottom right) on the staggered-grid
cell have remained unexcited (see text for further explanation). The four black circles delineate one velocity staggered-grid cell. The velocity fields are not
accurately computed due to the incomplete excitation of the pressure grid.

old system (x, z). This complementary approach has been developed in the time domain for the P-SV wave equation by Saenger et al.
(2000).
Equation system (3) becomes:
√  
−iωξx (x) 2 ∂ Q(x, z, t) ∂ Q(x, z, t)
Px (x, z, ω) = + + S(x, z, t)
K (x, z) 2 ∂x ∂z
√  
−iωξz (z) 2 ∂ R(x, z, t) ∂ R(x, z, t)
Pz (x, z, ω) = − +
K (x, z) 2 ∂x ∂z
√  
2 b(x, z) ∂ P(x, z, t) ∂ P(x, z, t)
−iωQ(x, z, ω) = +
2 ξx (x) ∂x ∂z
√  
2 b(x, z) ∂ P(x, z, t) ∂ P(x, z, t) (11)
−iω R(x, z, ω) = − + .
2 ξz (z) ∂x ∂z


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1275

Equation system (11) is discretized using second-order centred staggered-grid stencils:

 
∂Q 1 1
≈ √ (Q i+1/2, j−1/2 − Q i−1/2, j+1/2 )
∂ x i, j 2
 
∂Q 1 1
≈ √ (Q i+1/2, j+1/2 − Q i−1/2, j−1/2 ).
∂z i, j 2 (12)
The configuration of this second staggered grid is shown in Fig. 1. The two particle velocity components are defined at the same node. These
nodes are different from the nodes of the first staggered grid. In contrast, the pressure is defined at the same nodes as for the first staggered
grid. Using the same procedure as for the classical Cartesian coordinate system, we obtain the following discretized wave equation:
     
−ω2 1 1 1 1 1
+ bi+1/2, j−1/2 + + bi−1/2, j+1/2 +
K i, j 42 ξxi ξxi+1/2 ξzi ξz j−1/2 ξxi ξxi−1/2 ξzi ξz j+1/2
   
1 1 1 1
+ bi+1/2, j+1/2 + + bi−1/2, j−1/2 + Pi, j
ξxi ξxi+1/2 ξzi ξz j+1/2 ξxi ξxi−1/2 ξzi ξz j−1/2
     
1 1 1 1 1
+ −P i+1, j bi+1/2, j−1/2 − + b i+1/2, j+1/2 −
42 ξxi ξxi+1/2 ξzi ξz j−1/2 ξxi ξxi+1/2 ξzi ξz j+1/2
    
1 1 1 1
− Pi−1, j bi−1/2, j+1/2 − + bi−1/2, j+1/2 −
ξxi ξxi−1/2 ξzi ξz j+1/2 ξxi ξxi−1/2 ξzi ξz j+1/2
    
1 1 1 1
+ Pi, j+1 bi−1/2, j+1/2 − + bi+1/2, j+1/2 −
ξxi ξxi−1/2 ξzi ξz j+1/2 ξxi ξxi+1/2 ξzi ξz j+1/2
    
1 1 1 1
+ Pi, j−1 bi+1/2, j−1/2 − + bi−1/2, j−1/2 −
ξxi ξxi+1/2 ξzi ξz j−1/2 ξxi ξxi−1/2 ξzi ξz j−1/2
 
1 1
− Pi+1, j−1 bi+1/2, j−1/2 +
ξxi ξxi+1/2 ξzi ξz j+1/2
 
1 1
− Pi−1, j+1 bi−1/2, j+1/2 +
ξxi ξxi−1/2 ξzi ξz j+1/2
 
1 1
− Pi+1, j+1 bi+1/2, j+1/2 +
ξxi ξxi+1/2 ξzi ξz j+1/2
 
1 1 (13)
− Pi−1, j−1 bi−1/2, j−1/2 + = Si, j .
ξxi ξxi−1/2 ξzi ξz j−1/2

This system of equations defines a nine-point stencil inside the whole domain. If we now consider gridpoints outside the PML layers
(i.e. ξ x = ξ z = 1) and regroup terms with respect to buoyancy, the equation system (13) reduces to
−ω2 1
Pi, j = [bi+1/2, j−1/2 (Pi+1, j−1 − Pi, j ) + bi−1/2, j+1/2 (Pi−1, j+1 − Pi, j )
K i, j 22
+ bi+1/2, j+1/2 (Pi+1, j+1 − Pi, j ) + bi−1/2, j−1/2 (Pi, j − Pi−1, j−1 )] + Si, j . (14)
Outside PML zones, eq. (14) defines a five-point rotated stencil of coefficients:
ω2 1
C1 = + (bi+1/2, j−1/2 + bi−1/2, j+1/2 + bi+1/2, j+1/2 + bi−1/2, j−1/2 )
K i, j 22
1
R2 = bi−1/2, j
22
1
R3 = bi+1/2, j
22
1
R4 = bi, j−1/2
22
1 (15)
R5 = bi, j+1/2 ,
22
with a geometrical distribution given in Fig. 1. This rotated staggered-grid five-point stencil differs from the previous one based on the classical
unrotated staggered-grid strategy. If one considers a homogeneous density, the stencil of the system of eq. (15) is identical to the stencil of
Jo et al. (1996) when using a = 0; i.e. considering only the discretization of the Laplacian term on the rotated coordinate system (see eq. 3
in Jo et al. 1996). The approach of Jo et al. (1996) was extended to the case of heterogeneous density models by Štekl (1997). The stencil
developed by Štekl (1997) when using a = 0 is identical to the one of the system of eq. (15).
As for the first staggered-grid geometry (see the end of the section 2.1) we perform wave propagation over a single non-staggered
√ grid
√
with a mesh spacing  ns , where variables P, Q and R are defined at each node using a centred√ finite-different stencil (−1/2 2,
√ 0, 1/2 2)
for the first-derivative operator approximated on the rotated grid. The coefficients ±1/2 2 account for the spatial support 2 2ns of the
centred stencil, used to compute the first-order derivative along the bisectrices. Again, the impulsional source is excited at one grid node. The


C 2004 RAS, GJI, 157, 1269–1296
1276 B. Hustedt, S. Operto and J. Virieux

pressure snapshot shows that, again, one subgrid stays strictly unexcited when considering a point source (Fig. 2b). Nevertheless, we observed
a difference compared with the first staggered-grid case. One can note in Fig. 2(b) that only one point out of five was excited on the full grid
(i.e. a non-staggered grid) while the rotated staggered-grid structure shown in Fig. 1 would suggest that one gridpoint out of two should have
been excited. The white square in Fig. 2b delineates the elementary staggered-grid pressure cell associated with the rotated staggered-grid
stencil (Fig. 1). One can note that two of the four nodes at the corners of the cell (the upper left and the bottom right) have remained unexcited
(Fig. 2b), contrary to the case of the classical staggered-grid stencil for which the four nodes have been excited (Fig. 2a). The fact that half of
the pressure staggered-grid nodes remained unexcited implies that the spatial derivatives of the pressure fields, and hence the velocity fields,
are not correctly computed. Indeed, both the horizontal and vertical differential operator stencils associated with the rotated staggered-grid
strategy involve the four points surrounding the central node where the derivative has to be estimated (see eqs 8, 9, 10 and 11 in Saenger et al.
2000). Two of these four nodes remained unexcited during the simulation leading to an erroneous estimation of the spatial partial derivatives
of the pressure field. As an illustration, the black circles in Fig. 2(b) delineate a staggered cell for velocity fields. The upper left velocity node
is computed from the surrounding four pressure nodes indicated by the white square—two of them are zero.
This behaviour of the rotated staggered-grid stencil (namely, the incomplete excitation of the pressure staggered grid when an impulsional
point source is considered) suggests that its dispersion properties are less good that of the classical staggered-grid stencil. Note, however, that
a simple way to remove this artefact is to strengthen the grid coupling by smoothing the source spatial distribution, namely by spreading the
point source excitation on several adjacent
√ nodes (see also the section entitled ‘Source excitation in the wavelet domain’ in Operto et al. 2002).
Moreover, the space increment of 2 2ns along the rotated direction of spatial differentiations is greater that of the classical staggered-grid
stencil for which a space increment of 2 ns is considered along the unrotated direction of spatial differentiations. A comparable accuracy of
the classical and the rotated staggered-grid stencils is reached when the space increment along the direction of spatial differentiations
√ is the
same (see Fig. 4 in Saenger et al. 2000, in the case of the elastic wave equation). This requires to use a grid interval of / 2 when using the
rotated staggered-grid stencil instead of  for the classical staggered grid.

3 I M P R O V E D F I N I T E - D I F F E R E N C E D I S C R E T I Z AT I O N

How may we increase the accuracy of our numerical operators while maintaining computational efficiency? One strategy is to augment the
accuracy by incorporating gridpoints in the FD stencil around the collocation along several directions (Jo et al. 1996; Štekl & Pratt 1998; Shin
& Sohn 1998). This leads to the so-called mixed-grid stencils. Another approach consists of increasing the accuracy of FD operators along
Cartesian axes associated with the derivative directions (either the classical or rotated Cartesian axes) by increasing the order of the operator.

3.1 The mixed-grid stencil

When considering the standard five-point stencil, numerical dispersion is most significant along bisectrices. This dispersion is reduced by
adding four additional points to the five-point stencil. Jo et al. (1996) and Štekl (1997) have combined the acoustic wave equations written in
the original Cartesian coordinate system (eq. 10) and the 45◦ -rotated coordinate system (eq. 11). The two coordinate systems are discretized,
such
√ that they define the same numerical grid for the pressure field. This implies that the mesh spacing in the rotated coordinate system is equal
to 2. Combination of the wave equations in the two coordinate systems leads to a mixed-grid stencil that incorporates nine points (Fig. 3).
In the acoustic case, this combination requires only the development of the derivative approximations on the Cartesian and the 45◦ -rotated
grids, since the pressure is a scalar field. In the elastic case, additional algebra is required to manipulate vector fields (see Štekl & Pratt 1998,
for details). The mixed-operator approach is a state-of-the-art tool in FDFD seismic wave modelling for efficient and accurate modelling (Jo
et al. 1996; Štekl & Pratt 1998; Shin & Sohn 1998).
Discretization of the spatial derivatives on the 45◦ -rotated grid gives the following stencil formulation:
    
∂ 1 ∂ 1 1 1
P(x , z , ω) ≈ (P i+1, j−1 − P i, j ) − (P i, j − P i−1, j+1 )
∂ x ρ(x , z ) ∂ x i, j 22 ρi+1/2, j−1/2 ρi−1/2, j+1/2
    
∂ 1 ∂ 1 1 1
P(x , z , ω) ≈ (P i+1, j+1 − Pi, j ) − (Pi, j − Pi−1, j−1 ) , (16)
∂z ρ(x , z ) ∂z i, j 22 ρi+1/2, j+1/2 ρi−1/2, j−1/2
which is identical to the parsimonious rotated staggered-grid formulation described previously. We denote differential operators in the following
compact form:
     
∂ 1 ∂ ∂ 1 ∂
i, j = P(x, z, ω) + P(x, z, ω)
∂ x ρ(x, z) ∂ x i, j ∂z ρ(x, z) ∂z i, j
     
∂ 1 ∂ ∂ 1 ∂
i, j = P(x , z , ω) + P(x , z , ω) . (17)
∂ x ρ(x , z ) ∂ x i, j ∂z ρ(x , z ) ∂z i, j

We combine operators to an improved discrete wave equation that defines the mixed-grid stencil:
 
ω2
a i, j + (1 − a) i, j + P(x, z, ω) = Si, j . (18)
K (x, z) i, j


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1277

Figure 3. Illustration of three different stencils used for the finite-difference discretization of the acoustic wave equation. The nine-point mixed-grid stencil
and the 13-point staggered-grid stencil are considered for approximations of fourth-order accuracy. Another nine-point stencil, called the nine-point Cartesian
stencil, can be developed for a homogeneous medium when considering quantities along Cartesian axes.

The averaging coefficient a accounts for the coupling between grid nodes that lie on the Cartesian grid (black dots) and the rotated grid
(grey dots) (Fig. 1).
This analysis demonstrates the connection between staggered-grid and mixed-grid formulations. The mixed operator approach linearly
combines two staggered-grid formulations based on a second-order centred approximation of the first derivative in the hyperbolic system of
eqs (1) and (11). As we have shown, the two staggered-grid stencils lead to different five-point stencils. Combination of the two stencils gives
a nine-point stencil that has a very compact geometry (Fig. 3). The two staggered-grid systems do not have the same geometry because the
particle velocities do not follow the same configuration on the two staggered-grid formulations (Fig. 1). This also implies that the mixed-grid
stencil cannot be implemented in the first-order hyperbolic formulation using staggered grids (i.e. without adding additional intermediate grid
nodes). Only the parsimonious formulation can be implemented since it eliminates those wavefields for which the gridpoints are not located
on the same numerical grids, i.e. the particle velocity wavefields.
Does this discrete construction of derivative stencils avoid the decoupling of subgrid structures? Again, we performed wave propagation
simulations using the mixed-grid stencil for a point source in the same homogeneous medium we have considered for the staggered-grid
stencil simulations. The mixed-grid formulation (eq. 18) is recast into the framework of the first-order hyperbolic formulation, which gives:
    
∂ P(x, z, t) ∂ Q(x, z, t) ∂ R(x, z, t) ∂ Q (x , z , t) ∂ R (x , z , t)
= K (x, z) a + + (1 − a) + + S(x, z, t) (19)
∂t ∂x ∂z ∂x ∂z

∂ Q(x, z, t) ∂ P(x, z, t)
= b(x, z)
∂t ∂x
∂ R(x, z, t) ∂ P(x, z, t)
= b(x, z)
∂t ∂z
∂ Q (x , z , t) ∂ P(x , z , t)
= b(x , z )
∂t ∂x
∂ R (x , z , t) ∂ P(x , z , t)
= b(x , z ) , (20)
∂t ∂z
where Q and R are the acoustic components of the particle velocity vector in the rotated coordinate system. Again, the first-derivative operator
is approximated by a centred FD stencil (−1/2, 0, 1/2) on the full grid.
The snapshot of the pressure field shows decoupled grid effects, i.e. an unexcited grid is still present (Fig. 4a). Therefore, combining
the Cartesian five-point stencil and the rotated five-point stencil does not couple all nodes of the full pressure grid. The pressure gridpoints
that are excited by the combined stencil are only coupled at their common staggered-grid locations (Fig. 1). The pressure-grid coupling is
foreseeable since the locations of the excited gridpoints of the rotated staggered-grid stencil (Fig. 2b) and the classical staggered-grid stencil
(Fig. 2a) coincide. Note, however, that twice as many gridpoints are excited by the classical staggered-grid stencil as the rotated one (compare
Figs 2a and 2b).


C 2004 RAS, GJI, 157, 1269–1296
1278 B. Hustedt, S. Operto and J. Virieux

Distance (m) Distance (m)

(a) 0 1000 2000 3000 4000 900 910 920 930 940 950 960 970 980 990 1000

0 1400

1410

1420
1000
1430

1440

Depth (m)
Depth (m)

2000 1450

1460

1470
3000
1480

1490

4000 1500

Distance (m) Distance (m)

(b) 0 1000 2000 3000 4000 900 910 920 930 940 950 960 970 980 990 1000
0 1400

1410

1420
1000
1430

1440
Depth (m)

Depth (m)
2000 1450

1460

1470
3000
1480

1490

4000 1500

Distance (m) Distance (m)

(c) 0 1000 2000 3000 4000 900 910 920 930 940 950 960 970 980 990 1000

0 1400

1410

1420
1000
1430

1440
Depth (m)

Depth (m)

2000 1450

1460

1470
3000
1480

1490

4000 1500

Figure 4. (a, left) Snapshot of the pressure field propagating in a homogeneous medium and excited by an impulsive point source. The mixed-grid stencil is
implemented in the first-order hyperbolic formulation of the acoustic wave equation without acceleration averaging: two uncoupled grids are observed. The
excited grid is the combination of the excited grids of Figs 2(a) and (b). (a, right) Close-up of the snapshot. The intersection between the dashed lines indicates
the location of the gridpoints. (b, left) Snapshot of the additional variable Q that is analogous to the horizontal velocity snapshot for the elastic wave propagation.
(b, right) Close-up. (c, left) Snapshot of the additional variable R: vertical velocity snapshot. (c, right) Close-up. The mixed-grid formulation partially couples
the two combined staggered-grid stencils. In fact, the combined pressure grid shown in (a) excites two different staggered velocity grids. Therefore, additional
coupling of differently excited grids by an acceleration averaging strategy is a key element of this stencil, not just a numerical improvement.

√ √
We plot the snapshots for the particle velocities in the Cartesian coordinate system, 0.5[Q + 22 (Q + R )] and 0.5[R + 22 (−Q + R ))],
in Fig. 4. We observe that the excited velocity grids for the two components have nodes at different locations. The coupling of the two grids is
ensured by the excitation of the pressure grid, as is shown in Fig. 4(a). The coupling of the two velocity grids partially fills up the unexcited
grids (Figs 4b and c) vertically and horizontally for the horizontal and vertical velocities respectively. The observed pattern implies that
the coupling of the two stencils is not optimal. Moreover, we deduce second-order accuracy in the space increment  for the mixed-grid
stencil.


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1279

(a) 1/G[Number of grid points per wavelength]

0.0 0.1 0.2 0.3
1.10

1.05

Vph/v 1.00

0.95

0.90

(b) 1/G[Number of grid points per wavelength]

0.0 0.1 0.2 0.3
1.10

1.05
Vgh/v

1.00

0.95

0.90

Figure 5. Relative phase v ph and group v gr velocity curves for the mixed-grid formulation. We use similar optimization coefficients to Jo et al. (1996): a =
0.5461, c = 0.6248, d = 0.9381 × 10−1 and e = 0.25[(1 − 4) (d − c)]. Propagation angles range from 0◦ to 45◦ by in steps of 15◦ from the vertical. The
curves are denoted by stars, triangles, circles and squares respectively.

It was already noted that the dispersion of the mixed-grid stencil is more significant than that of the second-order staggered-grid stencil,
in the classical coordinate system (compare Figs 2 and 5 in Jo et al. 1996). This is in agreement with the partial coupling of the velocity grids
in the mixed-grid stencil. The unfavourable coupling of grids degrades the accuracy of the stencil. Therefore, another feature is introduced in
the mixed-grid stencil in order to strengthen the subgrid couplings leading to an improvement in modelling accuracy.
In view of this, the pressure acceleration term ω2 /K(x, z) should not be taken as a term implying only the collocation point (the so-called
lump approximation): one may approximate it using a weighted average over the mixed operator stencil nodes, as is done in finite-element
modelling (Zienkiewicz & Morgan 1982). This averaging procedure couples the subgrids. In spirit, the procedure which consists of averaging
the mass acceleration term is close to the smoothing of the spatial distribution of the source suggested for FDTD modelling at the end of
section 2.3.
The operator is defined as:
  
ω2 ω2
P ⇒ c Pi, j + d(Pi+1, j + Pi−1, j + Pi, j+1 + Pi, j−1 )
K (x, z) i, j K (x, z)

1 − c − 4d
+ (Pi+1, j+1 + Pi−1, j+1 + Pi+1, j−1 + Pi−1, j−1 ) , (21)
4
where c and d are two additional averaging coefficients accounting for the coupling between the Cartesian and the 45◦ -rotated grid. Numerically
accurate results, up to four points per shortest wavelength, are guaranteed by an optimization technique that determines the averaging coefficients


C 2004 RAS, GJI, 157, 1269–1296
1280 B. Hustedt, S. Operto and J. Virieux

1/G[Number of grid points per wavelength]

(a)
0.0 0.1 0.2 0.3
1.10

1.05

Vph/v
1.00

0.95

0.90

1/G[Number of grid points per wavelength]

(b)
0.0 0.1 0.2 0.3
1.10

1.05

1.00
Vgh/v

0.95

0.90

0.85

0.80

Figure 6. Relative phase v ph and group v gr velocity curves for the staggered-grid formulation using fourth-order staggered-grid stencils (α 1 = 9/8 and
α 2 = −1/24). We use the same symbols for propagation angles as in Fig. 5.

such that the phase velocity dispersion is minimized. Similar optimizations have been done for the FDTD approach by Holberg (1987). The
optimal coefficients computed by Jo et al. (1996) are a = 0.5461, c = 0.6248 and d = 0.09381. The value for the coefficient 1 − c − 4d =
0.1297 × 10−5 is negligible (Štekl 1997) and leads to a five-point distribution of the pressure acceleration instead of nine points. Therefore,
the complete stencil involving nine points is described by the following factors (Fig. 3):
 
ω2 a 1 1 1 1
C1 = c − 2 + + +
K i, j  ρi+1/2, j ρi−1/2, j ρi, j−1/2 ρi, j+1/2
 
1−a 1 1 1 1
− + + +
22 ρi+1/2, j−1/2 ρi−1/2, j+1/2 ρi+1/2, j+1/2 ρi−1/2, j−1/2
ω2 a 1
C2 = c + 2
K i, j  ρi−1/2, j
ω2 a 1
C3 = c + 2
K i, j  ρi+1/2, j
ω2 a 1
C4 = c + 2
K i, j  ρi, j−1/2
ω2 a 1
C5 = c + 2
K i, j  ρi, j+1/2


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1281

(a) 1/G[Number of grid points per wavelength]

0.0 0.1 0.2 0.3
1.10

1.05

Vph/v 1.00

0.95

0.90

(b) 1/G[Number of grid points per wavelength]

0.0 0.1 0.2 0.3
1.10

1.05

1.00
Vgh/v

0.95

0.90

0.85

0.80

Figure 7. Relative phase v ph and group v gr velocity curves for the optimized staggered-grid stencil. The operator coefficients used for the numerical
differentiation are α 1 = 1.13824281853071 and α 2 = −4.64142728435701 × 10−02 . We use the same symbols for propagation angles as in Fig. 5.

1−a 1
R1 =
2 ρi−1/2, j−1/2
1−a 1
R2 =
 ρi+1/2, j−1/2
2

1−a 1
R3 =
2 ρi−1/2, j+1/2
1−a 1
R4 = .
2 ρi+1/2, j+1/2 (22)

Although partial derivatives are approximated by second-order finite-difference operators, we expect this scheme to behave as a fourth-
order operator with respect to accuracy, since both mixed-grid and fourth-order staggered-grid stencils require comparable discretization rules
(i.e. five gridpoints per wavelength) (Levander 1988; Jo et al. 1996). Moreover, the numerical efficiency should be optimal because it is a
compact stencil. Numerical anisotropy is reduced as suggested by the stencil geometry (Fig. 3). Nevertheless, a more quantitative comparison
should be given by a dispersion analysis. We discuss this after we have accomplished the stencil construction.
One must keep in mind that the acceleration averaging technique improves the modelling accuracy for a homogeneous medium. Since
averaging is a smoothing technique it may pollute the solution in the presence of sharp boundaries. Let us also underline that the comparison
through the first-order hyperbolic formulation leads to an original implementation of a PML absorbing boundary condition for the two


C 2004 RAS, GJI, 157, 1269–1296
1282 B. Hustedt, S. Operto and J. Virieux

1/G[Number of grid points per wavelength]

(a)
0.0 0.1 0.2 0.3
1.10

1.05

Vph/v
1.00

0.95

0.90

1/G[Number of grid points per wavelength]

(b)
0.0 0.1 0.2 0.3
1.10

1.05
Vg/v

1.00

0.95

0.90

Figure 8. Relative phase v ph and group v gr velocity curves for the staggered-grid formulation using fourth-order staggered-grid stencils combined with the
acceleration averaging. We use the same symbols for propagation angles as in Fig. 5. The acceleration averaging significantly improves the phase velocity
dispersion curve (compare with Fig. 6).

staggered-grid stencils in the Cartesian and rotated coordinate system (eqs 7 and 13) and therefore for the mixed-grid stencil. An original
nine-point PML stencil has been constructed for the mixed-grid stencil, whilst only an approximate five-point stencil has been used previously
inside PML zones, as discussed by Dessa (2001) and Hustedt (2002).
In the numerical examples presented later on, we have combined the PML absorbing boundary condition with the 45◦ one-way paraxial
condition (condition A3 of eq. 10 in Clayton & Engquist 1977) at the outer edges of the model.

3.2 Fourth-order finite-difference discretization

Now we investigate a second type of approach to improve the accuracy of the acoustic wave modelling. We use fourth-order FD operators in
only a single Cartesian coordinate system.

3.2.1 The homogeneous Cartesian nine-point stencil

The first possible approach considers a fourth-order finite-difference approximation of the second derivative. A centred FD stencil along one
axis of coefficients (−1/12, 16/12, −30/12, 16/12, −1/12) can be deduced by classical Taylor expansion analysis (Nougier 1989). This leads
to a nine-point stencil for the Laplacian. The geometry of this stencil is shown in Fig. 3 (see the nine-point Cartesian stencil). This approach
is limited to homogeneous density models and is mentioned for purpose of completeness only. Therefore, it will not be investigated further.


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1283

1/G[Number of grid points per wavelength]

(a)
0.0 0.1 0.2 0.3
1.10

1.05

Vph/v
1.00

0.95

0.90
1/G[Number of grid points per wavelength]
(b) 0.0 0.1 0.2 0.3
1.10

1.05
Vg/v

1.00

0.95

0.90

Figure 9. Relative phase v ph and group v gr velocity curves for the optimized staggered-grid stencil combined with acceleration averaging. We use the same
symbols for propagation angles as in Fig. 5. The mass averaging significantly improves the phase velocity dispersion curve (compare with Fig. 7).

3.2.2 The fourth-order staggered-grid stencil

In order to account for heterogeneous density models, we come back to the first-order hyperbolic formulation and use fourth-order spatial
derivative approximations for the first derivative on staggered grids (Levander 1988).
   
∂Q 1 9 1
≈ (Q i+1/2, j − Q i−1/2, j ) − (Q i+3/2, j − Q i−3/2, j )
∂ x i, j  8 24
   
∂R 1 9 1 (23)
≈ (Ri, j+1/2 − Ri, j−1/2 ) − (Ri, j+3/2 − Ri, j−3/2 ) .
∂z i, j  8 24

The fourth-order staggered-grid stencil is the most commonly used in the time domain, since it is supposed to provide the best trade-off
between numerical accuracy and computational efficiency.
In the following, we develop the fourth-order staggered-grid stencil in the classical Cartesian coordinate system. A similar derivation
could be developed in the rotated frame by simply considering fourth-order finite-difference formulae along the 45◦ -rotated axes. These higher-
order schemes are deduced from the work of Saenger et al. (2000) and have been used by Cruz-Atienza & Virieux (2003) for time-domain
crack modelling.
The resulting stencil for the second-order wave equation incorporates 13 grid points (details are given in Appendix A). The staggered-grid
stencil is plotted in Fig. 3. We apply a similar derivation procedure as the one for the case of the second-order derivation approximation, and


C 2004 RAS, GJI, 157, 1269–1296
1284 B. Hustedt, S. Operto and J. Virieux

obtain for the pressure wavefield P:

   
−ω2 1 9 bi+1/2, j 9 1
Pi, j = 2 (Pi+1, j − Pi, j ) − (Pi+2, j − Pi−1, j )
K i, j  ξi 8 ξxi+1/2 8 24
 
bi−1/2, j 9 1
− (Pi, j − Pi−1, j ) − (Pi+1, j − Pi−2, j )
ξxi−1/2 8 24
  
1 bi+3/2, j 9 1
− (Pi+2, j − Pi+1, j ) − (Pi+3, j − Pi, j )
24 ξxi+3/2 8 24
  
bi−3/2, j 9 1
− (Pi−1, j − Pi−2, j ) − (Pi, j − Pi−3, j )
ξxi−3/2 8 24
   
1 9 bi, j+1/2 9 1
+ 2 (Pi, j+1 − Pi, j ) − (Pi, j+2 − Pi, j−1 )
 ξ j 8 ξz j+1/2 8 24
 
bi, j−1/2 9 1
− (Pi, j − Pi, j−1 ) − (Pi, j+1 − Pi, j−2 )
ξz j−1/2 8 24
  
1 bi, j+3/2 9 1
− (Pi, j+2 − Pi, j+1 ) − (Pi, j+3 − Pi, j )
24 ξz j+3/2 8 24
  
bi, j−3/2 9 1
− (Pi, j−1 − Pi, j−2 ) − (Pi, j − Pi, j−3 ) + Si, j . (24)
ξz j−3/2 8 24

The pressure acceleration term ω2 /K(x, z) can also be averaged over the different nodes of the FD stencil as for the mixed-grid strategy.
In this case eq. (24) becomes:
 
−ω2 −ω2 −ω2 −ω2 −ω2
astg Pi, j + cstg Pi+1, j + Pi−1, j + Pi−1, j−1 + Pi, j+1
K i, j K i+1, j K i−1, j K i, j−1 K i, j+1
 
−ω 2
−ω 2
−ω 2
−ω2
+ dstg Pi−2, j + Pi+2, j + Pi−1, j−2 + Pi, j+2
K i−2, j K i+2, j K i, j−2 K i, j+2
 
−ω 2
−ω 2
−ω 2
−ω2
+ estg Pi−3, j + Pi+3, j + Pi, j−3 + Pi, j+3
K i−3, j K i+3, j K i, j−3 K i, j+3
   
1 9 bi+1/2, j 9 1
= 2 (Pi+1, j − Pi, j ) − (Pi+2, j − Pi−1, j )
 ξi 8 ξxi+1/2 8 24
 
bi−1/2, j 9 1
− (Pi, j − Pi−1, j ) − (Pi+1, j − Pi−2, j )
ξxi−1/2 8 24
  
1 bi+3/2, j 9 1
− (Pi+2, j − Pi+1, j ) − (Pi+3, j − Pi, j )
24 ξxi+3/2 8 24
  
bi−3/2, j 9 1
− (Pi−1, j − Pi−2, j ) − (Pi, j − Pi−3, j )
ξxi−3/2 8 24
   
1 9 1 9 1
+ 2 (Pi, j+1 − Pi, j ) − (Pi, j+2 − Pi, j−1 )
 ξ j 8 ξz j+1/2 ρi, j+1/2 8 24
 
bi, j−1/2 9 1
− (Pi, j − Pi, j−1 ) − (Pi, j+1 − Pi, j−2 )
ξz j−1/2 8 24
  
1 bi, j+3/2 9 1
− (Pi, j+2 − Pi, j+1 ) − (Pi, j+3 − Pi, j )
24 ξz j+3/2 8 24
   (25)
bi, j−3/2 9 1
− (Pi, j−1 − Pi, j−2 ) − (Pi, j − Pi, j−3 ) + Si, j ,
ξz j−3/2 8 24
where e stg = [(1 − a stg )/4] − c stg − d stg . The coefficients a stg , c stg and d stg were computed such that the phase velocity dispersion is minimized
in a least-square sense, following the approach used by Jo et al. (1996) and Štekl & Pratt (1998). The final coefficients that we obtained are
a stg = 0.788614, b stg = 7.92484 × 10−2 and c stg = −3.22465 × 10−2 .
Let us underline that the averaging procedure is not essential to the unrotated staggered-grid numerical scheme. This is supported by
the time-domain modelling shown in Fig. 2(a) which showed that the correct pressure field was computed at the staggered-grid nodes for
a point source excited at one single grid node (i.e. without any smoothing of the source spatial distribution). In contrast, the erroneous
pressure field is computed by the rotated numerical scheme when a point source is excited at one single grid node (Fig. 2b). This suggests that


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1285

(a)

(b)

Figure 10. Illustration of the structure of the complex impedance matrix. Note that only the real part is shown. In (a) we display the result from the grid rotation
approach. The matrix contains (nx × nz)2 elements from which 9 × nx × nz are non-zero. In (b) we plot the matrix structure that results from a fourth-order
staggered-grid discretization of the first-order hyperbolic acoustic wave equation, reformulated to the parsimonious second-order hyperbolic equation. The
matrix contains (nx × nz)2 elements from which 13 × nx × nz are non-zero. Note that in both (a) and (b) the main diagonal is plotted with a thicker line.

any numerical scheme involving the rotated-grid strategy (i.e. either the rotated staggered-grid or the mixed-grid strategies) requires some
smoothing procedure such as the acceleration averaging technique or smoothing of the source spatial distribution.
In the time domain, acceleration averaging leads to an implicit formulation of the time integration, which has only been considered by
few authors (e.g. Mufti 1985) in seismology. The main drawback is the induced matrix inversion, which may limit the size of the domain to be
considered. One advantage is the possibility to select a coarser time stepping, which could be critical when different grid sizes are considered.
In the frequency domain, the matrix inversion is required in any case, and consequently the averaging procedure can be performed at
nearly no additional cost and should therefore be selected even for this staggered-grid formulation.

4 G R I D D I S P E R S I O N A N A LY S I S

In this section, numerical precision of the fourth-order staggered-grid stencil is investigated for both the wavelength content and azimuthal
variation. Furthermore, the staggered-grid approach is compared with the results obtained by Jo et al. (1996) for the mixed-grid approach. In
Figs 5 and 6, we show dispersion curves for relative phase and group velocity for both the mixed-grid stencil and the fourth-order staggered-
grid stencil, for plane wave propagation angles ranging from 0◦ to 45◦ . The number of gridpoints per wavelength is defined by 1/G(G is the
number of gridpoints per wavelength). In Fig. 6, we used non-optimized fourth-order staggered-grid derivative approximations for the stencil
and no averaging of the pressure acceleration term as done by Levander (1988).
From Figs 5 and 6 we find that the mixed-grid stencil shows less dispersive wave propagation for the same number of grid nodes per
shortest wavelength compared with the staggered-grid stencil. This is true for both phase and group velocity dispersion curves.
A partly similar optimization strategy, first introduced for FDTD modelling by Holberg (1987), can be applied to the staggered-
grid approach (Arntsen et al. 1998). In their approach, coefficients describing classical staggered derivative operator coefficients are re-
placed by optimized staggered operator coefficients that are obtained through a dispersion relation minimization technique. The fourth-order
staggered-grid operator coefficients (α 1 = 9/8 and α 2 = −1/24) were replaced by α 1 = 1.13824281853071 and α 2 = −4.64142728435701
× 10−2 . An example of the staggered-grid dispersion curve optimization is shown in Fig. 7. Both phase and group velocity dispersion curves
are slightly improved when using optimized coefficients.


C 2004 RAS, GJI, 157, 1269–1296
1286 B. Hustedt, S. Operto and J. Virieux

Figure 11. (a) Comparison of time-domain seismograms computed with the mixed-grid operator frequency-domain method (dashed line) and an analytical
solution (solid line). (b) Difference between the two sets of seismograms.

Finally, we compute the dispersion curves for the fourth-order staggered-grid operator using the coefficients of Holberg (1987) and
Levander (1988), combined with the averaging of the pressure acceleration term (eq. 25). The averaging coefficients that we computed for
the optimized stencil are a stg = 0.817876, c stg = 7.04578 × 10−2 and d stg = −3.10661 × 10−2 . When using averaging of the pressure
acceleration term we obtain a significant improvement of the phase velocity dispersion curve, both in terms of numerical dispersion and
numerical anisotropy (Figs 8 and 9). The accuracy of the fourth-order staggered-grid operator using optimized coefficients combined with
mass averaging is slightly better than the mixed-grid stencil. For 1/G ranging between 0 and 0.3, the maximum error for the phase velocity
is 0.3 per cent and 1.2 per cent for the fourth-order staggered-grid and mixed-grid formulations respectively. Note, however, that the group
velocity dispersion properties have been degraded by the mass term averaging (compare Figs 7, 8, and 9). Nevertheless, we are usually
concerned by the modelling of non-dispersive waves. Therefore we considered that the phase velocity dispersion is the most significant
criterion to assess the accuracy of a FD stencil (as suggested by the fact that only the phase velocity dispersion is minimized in the approach
developed by Jo et al. 1996; Štekl & Pratt 1998).
In Appendix B, we demonstrate that the staggered-grid strategy exactly determines relative phase velocity v ph /v and group velocity
equal to 1 under the assumption of small propagation angles for an infinitely fine medium discretization (v ph /v = v gr /v = 1 for 1/G ⇒ 0).
The mixed-grid approach obtains approximately similar results for numerical phase and group velocity (v ph /v ≈ v gr /v ≈ 1), because
the numerical combination of the rotated grids (and the subsequent optimization of the numerical dispersion) shifts the dispersion curves.
Indeed, the dispersion curves are better aligned to the ideal relative medium velocity for higher 1/G ratios (coarser discretization) while, at
the same time, the accuracy for lower 1/G ratios is altered. In Fig. 6, this is observed from relative velocity values at 1/G → 0.
It is remarkable that the grid rotation preserves the very compact numerical stencil geometry while including the maximum amount of
existing information in the surroundings of the centre point—a clear advantage for numerical FD modelling in 2-D with mixed grids.

5 S O L U T I O N S F O R T H E M AT R I X E Q UAT I O N

These discrete frequency-domain wave equations (eqs 18 and 25) are recast into a matrix-type equation of the form:
M · p = s, (26)


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1287

Figure 12. (a) Comparison of time-domain seismograms computed with the fourth-order staggered-grid frequency-domain method (dashed line) and an
analytical solution (solid line). (b) Difference between the two sets of seismograms.

where M is the so-called impedance matrix, p is the pressure wavefield (desired solution) and s is the source function. The matrix M is
complex-valued and depends on the physical model parameters: the density ρ, the bulk modulus K, and the angular frequency ω.
We compute wavefield solutions for each source position and each frequency component by solving the corresponding matrix equation
(eq. 26) through an optimized minimum-order scheme developed by Davis & Duff (1997). The so-called UMFPACK solver combines a matrix
pre-ordering strategy, that limits the numerical fill-in during the LU decomposition, with an unsymmetric-pattern multifrontal factorization
scheme. The multifrontal approach decomposes the initial matrix M into a set of frontal submatrices, which allows the implementation of
optimized pivot strategies that stabilize and accelerate the matrix factorization process.
For each frequency component, the complex matrix can be constructed and is factorized into its LU factors. Pressure wavefield solutions
for each source position and frequency component are computed separately. Clearly, the LU decomposition needs to be computed only
once for each frequency component, making this approach very appealing especially for multisource experiments (Pratt & Worthington
1990).
For example, the complex impedance matrix construction and subsequent LU factorization for one frequency component f j requires
98 per cent CPU time (10 per cent matrix construction +88 per cent matrix factorization) while the solving phase (LU f j · p f j = sif j ) for one
source position si only needs 2 per cent CPU time. Note that if a simple band-solver were to have been used for the matrix decomposition, the
number of non-zero coefficients in the LU factors would be significantly increased and, therefore, solving phase computations would be much
slower. Only when highly optimized matrix factorization methods are employed, such as multifrontal methods or nested dissection (Štekl &
Pratt 1998), does the storage and operation count reduce dramatically.
In the case of the mixed-grid stencil, the full matrix M contains (nx × nz)2 elements of which (9 × nx × nz) are different from zero.
The non-zero elements are distributed on the diagonal and two adjacent bands. The bandwidth is approximately 2nz + 1. The structure of the
impedance matrix is illustrated in Fig. 10(a). In the case of fourth-order derivative approximations, the staggered-grid discretization gives a
matrix M that contains (nx × nz)2 elements from which 13 × nx × nz are non-zero (Fig. 3). The elements are distributed along the main
diagonal (seven coefficients) and six adjacent parallel bands. The structure of the matrix is shown in Fig. 10(b). The bandwidth of the matrix
is 6nz, which is approximately three times larger than the matrix bandwidth resulting from the mixed-grid approach. The matrix appearing in
the first-order PML formulation increases the number of non-zero coefficients for a given model configuration, since 13 points instead of nine


C 2004 RAS, GJI, 157, 1269–1296
1288 B. Hustedt, S. Operto and J. Virieux

Figure 13. (a) Comparison of time-domain seismograms computed with the fourth-order staggered-grid frequency-domain method including acceleration
averaging (dashed line) and an analytical solution (solid line). (b) Difference between the two sets of seismograms.

points are used for the derivative approximations. Therefore, increasing the matrix bandwidth and number of non-zero coefficients results in
higher memory demands for the matrix ordering scheme.
The matrix characteristics such as the bandwidth, the number and the actual value of non-zero matrix coefficients are the key elements
that control the performance of the LU factorization procedure. For example, if a simple band-ordering solver is used, then the entire matrix
bandwidth that initially holds only zero coefficients will be filled-in with non-zeros. Even though we apply a sophisticated ordering scheme,
we expect the staggered-grid stencil construction to be less efficient in terms of CPU time, because of the spatial extension mentioned in the
previous paragraph. We shall see this with numerical examples below.

6 NUMERICAL EXAMPLES

6.1 The homogeneous full space model

We first compare the mixed-grid approach with the fourth-order staggered-grid approach for an acoustic wave simulation in a 2-D homogeneous
medium. The velocity is 4000 m s−1 and the density is 2500 kg m−3 . The vertical and horizontal dimensions of the model are 4000 m and
8000 m, respectively. The spatial sampling is 40 m, which corresponds approximately to eight points per shortest wavelength. The source is
an impulsional point source located at (z = 500 m, x = 1000 m). The source is excited at one single gridpoint (i.e. the vector s in eq. (26) has
only one non-zero element at the location of the impulsional source). For the temporal source excitation, we used a Gaussian derivative S(t)
= −2 α (t−t 0 ) e−α(t−t0 ) with α = 200, t 0 = 0.3 s. The central and maximum frequencies of the source are approximately 3.15 and 12 Hz
2

respectively. The minimum wavelength is 200 m. Forty receivers are spread out over the whole model at a constant depth of 400 m and with
a horizontal spacing between receivers of 200 m.
We compute frequency-domain solutions for 308 frequency components, spanning a range from 0 to 15 Hz. The inverse Fourier transform
is applied to the frequency-domain solutions to obtain time-domain seismograms. The maximum time of the seismograms is set to 20 s, for
accurate frequency sampling. For both approaches, we set PML absorbing boundary conditions on all edges of the model using the same
damping conditions throughout. In Fig. 11 we show the comparison between seismograms computed for the mixed-grid strategy (dashed line)


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1289

X (km)
0 1 2 3 4 5 6 7

2
Z (km)

Figure 14. 10 Hz monochromatic wavefield computed in the corner-edge model. The horizontal interface is at a depth of 1.88 km. The horizontal and vertical
interface are at a distance and depth of 1.88 km, respectively. The white dashed line delineates the PML layers. The wavefield is computed using the fourth-order
staggered-grid frequency-domain method.

and the analytical solution (solid line). Seismograms are plotted with their true amplitude (no amplitude normalization was applied to the
analytical and FD seismograms).
The overall agreement is excellent except for the receiver located just above the source, where the amplitude of the FD seismogram is
overestimated. This mismatch can be mitigated by smoothing the spatial distribution of the impulsional source, although accuracy of the FD
seismograms is slightly degraded at larger offsets (not shown).
The same comparison is shown in Figs 12 and 13 but for FD seismograms computed with the fourth-order staggered-grid stencil, with
and without averaging of the mass acceleration term respectively. The coefficients of Levander in the FD stencil were used (α 1 = 9/8 and
α 2 = −1/24). The fit with the analytical solution is slightly better than for the mixed-grid operator FD seismograms.
As expected, the computational time required by the mixed-grid formulation is much lower than for the staggered-grid stencil. The
time required to produce the time-domain seismograms is 45 min and 5 h for the mixed-grid operator and the fourth-order staggered-grid
simulations respectively. The modelling was carried out on a Linux PC with 512 Mbytes of RAM.

6.2 The corner edge model

The corner edge model combines two homogeneous media separated by a horizontal and a vertical interface making a corner (see Fig. 14).
The velocity and density in the two spaces are 2000 m s−1 , 8000 m s−1 and 2000 kg m−3 , 2500 kg m−3 , respectively. The dimension of the
model is 6400 m × 6400 m and is discretized with a mesh spacing of 40 m (equal to around four points per shortest wavelength). All four
edges of the model are treated with PML absorbing boundary conditions, where the layer size is set to 20 gridpoints. Therefore, the overall
grid has 200 × 200 gridpoints. As for the homogeneous case, the maximum time of the seismograms is set to 20 s for accurate frequency
sampling. This leads to 308 frequencies in the 0–15 Hz frequency range.
The source position is at (z = 240 m, x = 4000 m). We use the same source as for the homogeneous model comparison. The seismograms
are recorded by 81 receiver positions aligned horizontally over the whole model, at a depth of 160 m.
We compare time-domain seismograms computed with a O(t 2 , x 4 ) time-space staggered-grid method (see Operto et al. (2002) for
examples of application of this code in the case of the SH-wave equation) and the frequency-domain methods based on the mixed-grid and
fourth-order staggered-grid stencils. In Figs 15, 16 and 17, we show the comparison between the seismograms computed with the time-space
method and the frequency-domain method. For the latter simulations we apply the mixed-grid stencil and the fourth-order staggered-grid


C 2004 RAS, GJI, 157, 1269–1296
1290 B. Hustedt, S. Operto and J. Virieux

Figure 15. (a) Comparison of a finite-difference time-domain fourth-order staggered-grid solution (solid line) and the finite-difference frequency-domain
rotated-grid operator solution (dashed line) for the 2-D corner edge simulation. (b) Difference between the two sets of seismograms.

stencil with and without acceleration averaging respectively. We observe on the seismograms the direct wave, a strong reflection from the
horizontal edge, and a weak diffraction excited by the corner.
The three frequency-domain solutions agree well with the time-domain method solution. No amplitude normalization is applied to the
seismograms. We note that the residual between the solutions computed with the time-domain and the mixed-grid operator methods exhibit
some dispersion. This illustrates that the two stencils rely on different approaches. On the contrary, residuals between the time-domain and
the staggered-grid frequency-domain methods do not exhibit such dispersion, which is easily explained by the fact that the spatial stencils are
identical in both methods.
The time required to produce the time-domain seismograms is 51 min and 7.5 h for the mixed-grid operator and the fourth-order
staggered-grid simulations respectively. The modelling was carried out on a Linux PC with 512 Mbytes of RAM.

7 P E R S P E C T I V E S A N D C O N C LU S I O N
We present two staggered-grid strategies for FDFD acoustic wave propagation modelling in 2-D media. The two approaches discretize the
first derivative operator on the Cartesian coordinate system and the 45◦ rotated system, using a second-order centred FD stencil. The second
approach is the frequency-domain analogue of the time-domain approach developed by Saenger et al. (2000) for the elastic wave equation.
Staggered-grid approaches discretize the wave equation written as a first-order hyperbolic system of equations including PML absorbing
boundary conditions. Parsimonious elimination followed by reinjection leads to a second-order hyperbolic equation which allows minimal
numerical implementation schemes for both approaches. The two staggered-grid geometries share the same nodes for the pressure field, while
acoustic equivalents of the particle velocities are distributed on different nodes.


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1291

Figure 16. (a) Comparison of a finite-difference time-domain fourth-order staggered-grid solution (solid line) and the finite-difference frequency-domain
fourth-order staggered-grid solution (dashed line) for the 2-D corner edge simulation. (b) Difference between the two sets of seismograms.

We show that the staggered-grid stencil on the Cartesian coordinate system is identical to the centred five-point stencil strategy used by
Pratt & Worthington (1990), when terms due to the PML absorbing boundary conditions are neglected.
We extend the staggered-grid method to fourth-order accurate FD approximation of the first derivative on the Cartesian coordinate
system. The fourth-order staggered-grid method is compared with the mixed-grid method. The mixed-grid method linearly combines the two
second-order staggered-grid stencils on the Cartesian coordinate system and the rotated system. The mixed-grid method is accurate only if
an averaging of the acceleration term is incorporated in the stencil. This averaging strengthens the coupling between the two staggered-grid
stencils of different accuracy and geometry.
Numerical dispersion is investigated for the staggered-grid and mixed-grid stencils. For the mixed-grid discretization, both phase and
group velocity show less numerical dispersion than for the staggered-grid discretization. Since the combination of the stencils in the mixed-grid
method requires some dispersion relation optimization technique, we apply an analogous optimization to the staggered-grid approach in order
to enhance the approximation of the first-derivative operator.
Though numerical dispersion decreases when compared with the standard fourth-order staggered-grid approximation, it is still not
competitive with respect to the mixed-grid formulation. In the end, we add an acceleration term averaging to the optimized staggered-grid
stencil. In this case, the staggered-grid stencil turns out to be slightly superior to the mixed-grid stencil with respect to the phase velocity
dispersion.
The CPU efficiency of the two discretization approaches is controlled by the matrix factorization scheme and, therefore, by the compactness
of the discrete stencil geometry. Fewer surrounding points involved in the stencil construction are more favourable. Since the staggered-grid


C 2004 RAS, GJI, 157, 1269–1296
1292 B. Hustedt, S. Operto and J. Virieux

Figure 17. (a) Comparison of a finite-difference time-domain fourth-order staggered-grid solution (solid line) and the finite-difference frequency-domain
fourth-order staggered-grid, including acceleration averaging, solution (dashed line) for the 2-D corner edge simulation. (b) Difference between the two sets of
seismograms.

stencil for fourth-order approximations holds 13 gridpoints instead of nine points for the mixed-grid stencil, the memory requirements and
CPU floating point operations for matrix factorizations significantly increase.
Altogether, the mixed-grid strategy shows clearly superior results for 2-D FDFD modelling in terms of memory requirements and CPU
efficiency. In view of the performance of the mixed-grid approach the question arises as to why, in FDTD modelling, the staggered-grid
approach predominates while in FDFD modelling the rotated-grid approach is preferred. The main reason seems to be that in the latter case the
necessary mass acceleration averaging (i.e. antilumping of the mass matrix) implies matrix inversion at each step of the integration, from the
time t till the time t + dt. In FDTD modelling, the compactness of the discrete stencil plays a less important role since no matrix factorization
is computed. Moreover, the mixed-grid approach cannot be implemented in the time domain with a first-order hyperbolic formulation based
on staggered grids because of the conflicting geometry of the two stencils. Only a non-staggered-grid approach can be implemented, which
requires a decrease of the mesh stepping by 2.
The combination of rotated grids and staggered grids has already been addressed by Saenger et al. (2000). However, in their approach
only the derivative approximations are described on rotated axis while the wave equation is left untouched. They obtained superior simulation
results for wave diffractions by a crack, but did not address a higher-order scheme in order to enhance the numerical performance of the FDTD
scheme. Therefore, we propose that the compact and optimized rotated-grid stencil should be investigated for FDTD modelling in complex
heterogeneous media. This work is in progress.
We present and compare different techniques for the construction of finite-difference schemes. We show that for acoustic wave simulation
in the frequency domain, the mixed-grid scheme leads to the most adapted stencil due to its CPU efficiency. During this investigation, we


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1293

have shown the link between the mixed-grid method and the staggered-grid method. As an additional result, a new and more accurate PML
absorbing boundary condition arise for the mixed-grid method.

AC K N OW L E D G M E N T S

We are thankful to D. Rodrigues for providing us with the fourth-order staggered-grid stencil optimization coefficients. We would like to
thank P. Amestoy for assistance with the optimized minimum degree ordering scheme. Support from IDRIS (Institut du Développement et
des Ressources en Informatique Scientifique) through project 21560 is acknowledged. The research leading to this paper was partly founded
by DASE/CEA through the PROSIS3D group. The reviews provided by G. Pratt and K. Olsen were extremely helpful and led to significant
improvements in the quality of the manuscript. This paper is contribution no 643 from UMR Geosciences Azur, CNRS, France.

REFERENCES Luo, Y. & Schuster, G., 1990. Parsimonious staggered grid finite differencing
of the wave equation, Geophys. Res. Lett., 17(2), 155–158.
Arntsen, B., Nebel, A. & Amundsen, L., 1998. Visco-acoustic finite- Madariaga, R., 1976. Dynamics of an expanding circular fault, Bull. seism.
difference modeling in the frequency domain, J. Seism. Explor., 7, 45–64. Soc. Am., 73, 765–780.
Berenger, J.P., 1994. A perfectly matched layer for the absorption of elec- Mufti, I.R., 1985. Seismic modeling in the implicit mode, Geophys.
tromagnetic waves, J. Comput. Phys., 114, 185–200. Prospect., 33, 619–656.
Boore, D.M., 1972. Finite-difference methods for seismic wave propaga- Nougier, J.P., 1989. Méthodes de Calcul Numérique. Masson, Paris.
tion in heterogeneous materials, in Methods in Computational Physics, Operto, S., Virieux, J., Hustedt, B. & Malfanti, F., 2002. Adaptive wavelet-
eds Bolt, B.A., Alder, B., Frenbach, S., Rotemberg, M., 11, pp. 1–36, based finite-difference modelling of SH-wave propagation, Geophys. J.
Academic Press, New York. Int., 148(3), 477–499.
Clayton, R. & Engquist, B., 1977. Absorbing boundary conditions for acous- Pratt, R.G., 1990. Inverse theory applied to multi-source cross-hole tomog-
tic and elastic wave equations, Bull. seism. Soc. Am., 67, 1529–1540. raphy, part 2: Elastic wave-equation method, Geophys. Prospect., 38, 311–
Cruz-Atienza, V.M. & Virieux, J., 2003. Dynamic rupture simulation of bent 329.
faults with a finite difference approach, Geophys. J. Int., in press. Pratt, R.G. & Worthington, M.H., 1990. Inverse theory applied to multi-
Dablain, M.A., 1986. The application of higher order differencing to the source cross-hole tomography, part 1: Acoustic wave-equation method,
scalar wave equation, Geophysics, 51, 54–66. Geophys. Prospect., 38, 287–310.
Davis, T.A. & Duff, I.S., 1997. An unsymmetric-pattern multifrontal method Saenger, E.H., Gold, N. & Shapiro, A., 2000. Modeling the propagation of
for LU factorization, SIAM J. Sci. Stat. Comput., 18, 140–158. elastic waves using a modified finite-difference grid, Wave Motion, 31,
Dessa, J.X., 2001. Modélisation et inversion de la forme d’onde en domaine 77–92.
fréquentiel: applications à la tomographie multi-offsets, PhD thesis, Uni- Shin, C. & Sohn, H., 1998. A frequency-space 2-d scalar wave extrapo-
versity Paris-Sud. lator using extended 25-point finite-difference operator, Geophysics, 63,
Graves, R.W., 1996. Simulating seismic wave propagation in 3d elastic media 289–296.
using staggered grid finite differences, Bull. seism. Soc. Am., 86, 1091– Virieux, J., 1984. SH wave propagation in heterogeneous media, velocity-
1106. stress finite difference method, Geophysics, 49, 1259–1266.
Holberg, O., 1987. Computational aspects of the choice of operator and Virieux, J., 1986. P-SV wave propagation in heterogeneous media, velocity-
sampling interval for numerical differentiation in large-scale simulation stress finite difference method, Geophysics, 51, 889–901.
of wave phenomena, Geophys. Prospect., 35, 629–655. Štekl, I., 1997. Frequency domain seismic forward modelling: a tool for
Hustedt, B., 2002. Propagation des ondes: approches espace et ondelettes., waveform inversion, PhD thesis, Department of Geology, Royal School
PhD thesis, University of Nice-Sophia-Antipolis. of Mines, Imperial College London.
Jo, C.H., Shin, C.S. & Suh, J.H., 1996. An optimal 9 point, finite dif- Štekl, I. & Pratt, R.G., 1998. Accurate viscoelastic modeling by frequency-
ference, frequency-space, 2-D wave extrapolator, Geophysics, 61, 529– domain finite differences using rotated operators, Geophysics, 63(5),
537. 1779–1794.
Kelly, K.R., Ward, R.W., Treitel, S. & Alford, R.M., 1976. Synthetic seis- Zhang, Y.-G. & Ballmann, J., 1997. Two techniques for the absorption of
mograms: a finite difference approach, Geophysics, 41, 2–27. elastic waves using an artificial transition layer, Wave Motion, 25, 15–33.
Levander, A.R., 1988. Fourth-order finite-difference P-SV seismograms, Zienkiewicz, O.C. & Morgan, K., 1982. Finite Elements and Approximation,
Geophysics, 53, 1425–1436. Johm Wiley, New York.

A P P E N D I X A : F O U RT H - O R D E R D I S C R E T I Z AT I O N O F T H E WAV E E Q UAT I O N

We show the discretization procedure of the first-order hyperbolic system of equations using the fourth-order spatially staggered stencil. The
procedure is similar to the one used for second-order derivative approximations on staggered grids. This formulation leads to a staggered-grid
stencil that propagates pressure wavefield solutions with fourth-order accuracy.
We discretize eq. (21) using fourth-order staggered-grid geometry employed by Levander (1988):
 
∂Q 1 9 1
≈ (Q i+1/2, j − Q i−1/2, j ) − (Q i+3/2, j − Q i−3/2, j )
∂x  8 24
 
∂R 1 9 1
≈ (Ri, j+1/2 − Ri, j−1/2 ) − (Ri, j+3/2 − Ri, j−3/2 )
∂x  8 24
 
∂P 1 9 1
≈ (Pi+1, j − Pi, j ) − (Pi+2, j − Pi−1, j )
∂x  8 24
 
∂P 1 9 1
≈ (Pi, j+1 − Pi, j ) − (Pi, j+2 − Pi, j−1 ) , (A1)
∂z  8 24


C 2004 RAS, GJI, 157, 1269–1296
1294 B. Hustedt, S. Operto and J. Virieux

where  is the mesh spacing in the uniform grid.

We obtain the following relations:

 
−iωξxi 1 9 1
Pxi, j = (Q i+1/2, j − Q i−1/2, j ) − (Q i+3/2, j − Q i−3/2, j ) + Si, j
K i, j  8 24
 
−iωξzi 1 9 1
Pzi, j = (Ri+1/2, j − Ri−1/2, j ) − (Ri+3/2, j − Ri−3/2, j )
K i, j  8 24
 
1 1 9 1
Q i+1/2, j = − (Pi+1, j − Pi, j ) − (Pi+2, j − Pi−1, j )
iωξi+1/2 ρi+1/2, j  8 24
 
1 1 9 1
Q i−1/2, j = − (Pi, j − Pi−1, j ) − (Pi+1, j − Pi−2, j )
iωξi−1/2 ρi−1/2, j  8 24
 
1 1 9 1
Q i+3/2, j = − (Pi+2, j − Pi+1, j ) − (Pi+3, j − Pi, j )
iωξi+3/2 ρi+3/2, j  8 24
 
1 1 9 1
Q i−3/2, j = − (Pi−1, j − Pi−2, j ) − (Pi, j − Pi−3, j )
iωξi−3/2 ρi−3/2, j  8 24
 
1 1 9 1
Ri, j+1/2 = − (Pi, j+1 − Pi, j ) − (Pi, j+2 − Pi, j−1 )
iωξ j+1/2 ρi, j+1/2  8 24
 
1 1 9 1
Ri, j−1/2 = − (Pi, j − Pi, j−1 ) − (Pi, j+1 − Pi, j−2 )
iωξ j−1/2 ρi, j−1/2  8 24
 
1 1 9 1
Ri, j+3/2 = − (Pi, j+2 − Pi, j+1 ) − (Pi, j+3 − Pi, j )
iωξ j+3/2 ρi, j+3/2  8 24
  (A2)
1 1 9 1
Ri, j−3/2 = − (Pi, j−1 − Pi, j−2 ) − (Pi, j − Pi, j−3 ) ,
iωξ j−3/2 ρi, j−3/2  8 24

where we use effective medium parameters and PML boundary conditions as before. We combine the equations and obtain for the pressure
wavefield:
   
−ω2 1 9 1 9 1
Pi, j = 2 (Pi+1, j − Pi, j ) − (Pi+2, j − Pi−1, j )
K i, j  ξi 8 ξxi+1/2 ρi+1/2, j 8 24
 
1 9 1
− (Pi, j − Pi−1, j ) − (Pi+1, j − Pi−2, j )
ξxi−1/2 ρi−1/2, j 8 24
  
1 1 9 1
− (Pi+2, j − Pi+1, j ) − (Pi+3, j − Pi, j )
24 ξxi+3/2 ρi+3/2, j 8 24
 
1 9 1
− (Pi−1, j − Pi−2, j ) − (Pi, j − Pi−3, j )
ξxi−3/2 ρi−3/2, j 8 24
   
1 9 1 9 1
+ 2 (Pi, j+1 − Pi, j ) − (Pi, j+2 − Pi, j−1 )
 ξ j 8 ξz j+1/2 ρi, j+1/2 8 24
 
1 9 1
− (Pi, j − Pi, j−1 ) − (Pi, j+1 − Pi, j−2 )
ξz j−1/2 ρi, j−1/2 8 24
  
1 1 9 1
− (Pi, j+2 − Pi, j+1 ) − (Pi, j+3 − Pi, j )
24 ξz j+3/2 ρi, j+3/2 8 24
 
1 9 1 (A3)
− (Pi, j−1 − Pi, j−2 ) − (Pi, j − Pi, j−3 ) + Si, j .
ξz j−3/2 ρi, j−3/2 8 24

Finally, we write gridpoint contributions of the new staggered-grid stencil for fourth-order derivative approximations that are valid for the
whole medium, including PML absorbing boundary conditions. We note that the staggered-grid approach results in a stencil that incorporates


C 2004 RAS, GJI, 157, 1269–1296
 Frequency-domain acoustic wave modelling 1295

13 grid nodes (see Fig. 3):

  
ω2 1 99 1 1
C1 = − 2 +
K i, j  ξi 8 8 ξi+1/2 ρi+1/2, j ξi−1/2 ρi−1/2, j
 
1 1 1 1
+ +
24 24 ξi+3/2 ρi+3/2, j ξi−3/2 ρi−3/2, j
  
1 99 1 1
− 2 +
 ξ j 8 8 ξ j+1/2 ρi, j+1/2 ξ j−1/2 ρi, j−1/2
 
1 1 1 1
+ +
24 24 ξ j+3/2 ρi, j+3/2 ξ j−3/2 ρi, j−3/2
 
1 1 99 1 9 1 1 1 9 1
C2 = + 2 + +
 ξi 8 8 ξi−1/2 ρi−1/2, j 8 24 ξi+1/2 ρi+1/2, j 24 8 ξi−3/2 ρi−3/2, j
 
1 1 99 1 9 1 1 1 9 1
C3 = + 2 + +
 ξi 8 8 ξi+1/2 ρi+1/2, j 8 24 ξi−1/2 ρi−1/2, j 24 8 ξi+3/2 ρi+3/2, j
 
1 1 99 1 9 1 1 1 9 1
C4 = + 2 + +
 ξ j 8 8 ξ j−1/2 ρi, j−1/2 8 24 ξ j+1/2 ρi, j+1/2 24 8 ξ j−3/2 ρi, j−3/2
 
1 1 99 1 9 1 1 1 9 1
C5 = + 2 + +
 ξ j 8 8 ξ j+1/2 ρi, j+1/2 8 24 ξ j−1/2 ρi, j−1/2 24 8 ξ j+3/2 ρi, j+3/2
 
1 1 9 1 1 1 9 1
C6 = − 2 +
 ξi 8 24 ξi−1/2 ρi−1/2, j 24 8 ξi−3/2 ρi−3/2, j
 
1 1 9 1 1 1 9 1
C7 = − 2 +
 ξi 8 24 ξi+1/2 ρi+1/2, j 24 8 ξi+3/2 ρi+3/2, j
 
1 1 9 1 1 1 9 1
C8 = − 2 +
 ξ j 8 24 ξ j−1/2 ρi, j−1/2 24 8 ξ j−3/2 ρi, j−3/2
 
1 1 9 1 1 1 9 1
C9 = − 2 +
 ξ j 8 24 ξ j+1/2 ρi, j+1/2 24 8 ξ j+3/2 ρi, j+3/2
 
1 1 1 1 1
C10 = + 2
 ξi 24 24 ξi−3/2 ρi−3/2, j
 
1 1 1 1 1
C11 = + 2
 ξi 24 24 ξi+3/2 ρi+3/2, j
 
1 1 1 1 1
C12 = + 2
 ξ j 24 24 ξ j−3/2 ρi, j−3/2
 
1 1 1 1 1 (A4)
C13 = + 2 .
 ξ j 24 24 ξ j+3/2 ρi, j+3/2

A P P E N D I X B : D I S P E R S I O N A N A LY S I S

We substitute a plane wave solution of the form P = P 0 e−i(kx x+kz z) into eq. (3) for the fourth-order staggered-grid stencil. We further assume
x = z =  to be the uniform grid spacing in the x and the z directions. The propagation angle from the z-axis is denoted by θ , the spatial
wavenumber by k, and the operator coefficients for the staggered derivative approximation by α 1 and α 2 (for example, α 1 = 9/8 and α 1 =
−1/24 in the case of the fourth-order staggered-grid operator). In a discrete form, we have:
Pm,n = P0 · e−ik(m sin θ+n cos θ) . (B1)
Substitution of the discrete pressure field (eq. B1) in the staggered-grid wave eq. (eq. 24) gives the dispersion equation:
v2
ω2 (k) = [4α1 α1 − 2α1 α1 cos(k sin θ ) + 4α1 α2 cos(k sin θ)
2
− 4α1 α2 cos(2k sin θ ) − 2α2 α2 cos(3k sin θ)

+ 4α2 α2 − 2α1 α1 cos(k cos θ ) + 4α1 α2 cos(k cos θ)

(B2)
− 4α1 α2 cos(2k cos θ ) − 2α2 α2 cos(3k cos θ)] .


C 2004 RAS, GJI, 157, 1269–1296
1296 B. Hustedt, S. Operto and J. Virieux

We define the numerical phase and group velocities as v ph = ω/k and v gr = ∂ω/∂k respectively. From eq. (B2) we get the relative numerical
phase and group velocities:
vph
2
G2  π π π π
= 4α1 α1 sin2 sin θ + 8α1 α2 sin2 sin θ − 8α1 α2 sin2 2 sin θ + 4α2 α2 sin2 3 sin θ
v2 4π 2 G G G G
π π π π
+ 4α1 α1 sin2 cos θ + 8α1 α2 sin2 cos θ − 8α1 α2 sin2 2 cos θ + 4α2 α2 sin 2 3 cos θ (B3)
G G G G
∂ω 1
= {2α1 α1 [sin θ sin(k sin θ ) + cos θ sin(k cos θ )]
∂k 2
+ 4α1 α2 [sin θ sin(k sin θ ) + cos θ sin(k cos θ )]
− 8α1 α2 [sin θ sin(2k sin θ ) + cos θ sin(2k cos θ)]

+ 6α2 α2 [sin θ sin(3k sin θ ) + cos θ sin(3k cos θ)]} J −1/2 , (B4)
where G = λ/ = 2 π/(k/) denotes the number of grid points per wavelength λ and J is:
J = −2α1 α1 [cos(k sin θ ) + cos(k cos θ )] − 4α1 α2 [cos(k sin θ) + cos(k cos θ)]
+ 4α1 α2 [cos(2k sin θ ) + cos(2k cos θ )] − 2α2 α2 [cos(3k sin θ) + cos(3k cos θ)]

+ 4α1 α1 + 4α2 α2 . (B5)

In order to illustrate the determination of accurate wave velocities by the staggered-grid approach we assume sin θ ≈ θ for small angles θ
and simplify the phase velocity dispersion relations (eqs B2 and B4):
vph
= α1 − 3α2
v
vgr √
= α1 α1 − 6α1 α2 + 9α2 α2 . (B6)
v
Using fourth-order staggered operator coefficients α 1 = 9/8 and α 2 = −1/24 (Levander 1988) we deduce, for both phase and group velocity,
exactly 1.


C 2004 RAS, GJI, 157, 1269–1296