SPECTRAL WAVEFIELD EXTRAPOLATION METHODS

In imaging and inversion applications, and with the recent advances in our computa-
tional capacity, short changing the wave equation with high frequency approximations
is no longer necessary for modeling, and thus, waveeld time extrapolation methods
are now commonly used in practice (Geller and Takeuchi, 1998; Kristek et al., 2010).
Wave equation solutions based on waveeld time extrapolation are still relatively ex-
pensive. In addition, waveeld extrapolation methods, in general, have long been
accused of being big on results and short on the details, compared to those commonly
associated with waveeld geometrical methods (traveltimes).

Background

We typically, as we saw last week, use second-order, fourth-order, or higher-order

nite-dierence approximations of the wave equation to extrapolate waveelds. These
approximations provide reasonably ecient and exible extrapolation. They are also
very easy to implement and may include initial and boundary conditions, including
irregular boundary conditions. However, such an approach suers from dispersion
and stability related issues requiring ne space sampling, and consequently small
time extrapolation steps. It also does not allow for simple separation of wave modes
(solutions) necessary to isolate desired waveeld solutions like the P-wave mode in
anisotropic acoustic media.
On the other hand, spectral, including pseudo spectral and analytical, methods
for extrapolation have emerged recently to address some of these limitations (Reshef
et al., 1988a,b; Etgen and Brandsberg-Dahl, 2009; Chu and Stoa, 2011). Spectral
methods extrapolate waveelds by applying the Laplace operator as a phase shift
in the Fourier (wavenumber) domain. Thus, the extrapolation is overall dispersion
free and unconditionally stable for homogeneous media regardless of the length of
the time step. The Fourier implementation, however, poses challenges in handling
inhomogeneous media in an ecient way due to the mixed space-wavenumber domain
operator. These challenges are addressed using smart variable separation methods
(Etgen and Brandsberg-Dahl, 2009; Zhang and Zhang, 2009; Pestana and Stoa,
2010; Fomel et al., 2010), including simple Taylor's series type expansions of the
extrapolator operator (Etgen, 1989) to implement pseudo spectral methods. Despite
the eectiveness of these methods they are not cheap.
The waveeld extrapolation problem refers to advancement of a waveeld through
space or time. Both extrapolation in depth and extrapolation in time can be used
in seismic modeling and seismic migration. Reverse time migration, or RTM (Baysal
et al., 1983; McMechan, 1983; Whitmore, 1983; Levin, 1984), involves wave extrapo-
lation forward and backward in time. RTM is useful for accurate imaging in complex
areas and is drawing more and more attention as the most powerful depth-imaging
method (Yoon et al., 2004; Symes, 2007; Fletcher et al., 2009; Fowler et al., 2010).
Alkhalifah 2 Spectral methods

Reverse-time migration can correctly handle complex velocity models without dip
limitations on the image. However, it has large memory requirements and needs a
signicant amount of computation. The most popular and straightforward way to
implement reverse-time migration is the method of explicit nite dierences, which
is only conditionally stable because of the limit on time-step size. Finite-dierence
methods also suer from numerical dispersion problems, which can be overcome either
by decreasing the time step or by high-order schemes (Wu et al., 1996; Liu and Sen,
2009). Several alternative algorithms have been developed for seismic wave extrapola-
tion in variable velocity media. Soubaras and Zhang (2008) introduced an algorithm
based on a high-order dierential operator, which allows a large extrapolation time
step by solving a coecient optimization problem. Zhang and Zhang (2009) proposed
one-step extrapolation method by introducing a square-root operator. This method
can formulate the two-way wave equation as a rst-order partial dierential equation
in time similar to the one-way wave equation. Etgen and Brandsberg-Dahl (2009)
modied the Fourier Transform of the Laplacian operator to compensate exactly for
the error resulting from the second-order time marching scheme used in conventional
pseudo spectral methods (Reshef et al., 1988b). Fowler et al. (2010) provided an
accurate VTI P-wave modeling method with coupled second-order pseudo-acoustic
equations. Pestana and Stoa (2010) presented an application of Rapid Expansion
Method (REM) (Tal-Ezer et al., 1987) for forward modeling with one-step time evolu-
tion algorithm and RTM with recursive time stepping algorithm. As a chain operator
of Fast Fourier Transform and Finite Dierence operators, the proposed extrapola-
tor can be as accurate as the parameter interpolation approach employed by Etgen
and Brandsberg-Dahl (2009) but at a cost of only one Fast Fourier Transform (FFT)
and inverse Fast Fourier Transform (IFFT) operation. The advantages of the FFD
operator are even more apparent in the anisotropic case: no need for several inter-
polations for dierent parameters with the corresponding computational burden of
several FFTs and IFFTs. In addition, the operator can overcome the coupling of
qP-waves and qSV-waves (Zhang et al., 2009). We demonstrate the method on syn-
thetic examples and propose to incorporate FFD into reverse-time migration in order
to enhance migration accuracy and stability.

The spectral approach

Starting from the representation of the acoustic isotropic wave equation for homoge-
neous media in the Fourier wavenumber domain:
@ 2 P^
@t2
= jkj vP^ (k; t) (1)
@ 2 P^
@t2
= (k)P^ (k; t) (2)
where P^ (t; k) is the n-dimensional Fourier transform of P (t; x):
P (t; k) =
1 Z
(2 )n P (t; x)e dx ; (3)
i k
x
Alkhalifah 3 Spectral methods

and x represent the space coordinates x = fx; y; zg. The solution to this second-order
ordinary dierential equation is given by
P^ (k; t) = P^ (k; 0)ei (x;k)t : (4)
The  re ects two sets of ingoing and outgoing solutions. For a small enough time
step any inhomogeneous medium can be approximated to be locally homogeneous,
and thus
P^ (k; t +
t) = P^ (k; t)ei (x;k)
t : (5)
This is synonymous to Huygens principal for plane waves. However, to apply such
an operator in inhomogeneous media we need to access the space domain. Thus, for
inhomogeneous media, we consider a mixed-domain space-wavenumber operator, ,
for time stepping given by
Z
P (t +
t; x) = P (t; k) ei (x;k;
t) dk ; (6)
In the geometrical (high-frequency) approximation, the function (x; k; t) appearing
in equation 6 should satisfy the appropriate eikonal equation, which is, in the case of
isotropic 3D media, the dispersion relation
v
u !2
@ u 1
= v2(x; y; z) @t
t @
jrT j2; (7)
@z
where v(x; y; z) is the velocity of the medium.
In the case of VTI (vertical transverse isotropy) or TTI (tilted transverse isotropy),
the DSR equation can be eectively extended with the help of the acoustic approxi-
mation of Alkhalifah (1998). For simplicity, we limit the following discussion to the
isotropic 2D case.
If we consider equation 6 as an evolution in time, the initial condition for it is
(x; k; 0) = k
x ; (8)
which simply turns equation 6 into an inverse Fourier transform from wavenumber to
space coordinates.
For small steps
t, it is appropriate to replace  with its Taylor series approxi-
mation
(x; k; t)  k
x + 0 (x; k) t : (9)
Our nal approximation for the wave extrapolation operator, valid for small
t,
is Z
P (t 
t; x)  P (t; k) ei  (x;k)
t+i k
x dk
0
(10)
For a second order accuracy in time, we sum up the contributions of P (t +
t; x)
and P (t
t; x) to obtain:
Z
P (t +
t; x) + P (t
t; x)  2 P (t; k) cos [0 (x; k);
t] ei k
x dk : (11)
Alkhalifah 4 Spectral methods

Thus,
Z
P (t +
t; x)  2P (t; x) P (t
t; x) + 2 P (t; k) (cos [0(x; k);
t] 1) ei k
x dk :(12)
Equation 11 is too expensive to implement directly. One can, however, further
approximate it with nite-dierence or Fourier-domain operators.

The pseudo spectral approach

We can tie the waveeld extrapolation form 11 to the conventional second order space
domain representation we are accustom to and studied last week. This is done by
expanding the cosine function using Taylor series as follows:
cos [0(x; k)
t] = 1
t2 0 +
24 t4 40
t6 60
2 2  
720 + O
7 : (13)
Using only two terms of the expansion in equation 11, we obtain
Z
P (t +
t; x)  2P (t; x) P (t
t; x) + 2 P (t; k)

t220 ei k
x dk : (14)
2
For acoustic isotropic media, 0 = v(x) jkj, and thus,
P (t +
t; x)  2P (t; x) P (t
t; x) + v 2 (x)(
t)2 r2 P (t; x): (15)
This is our conventional wave equation with the 2nd order time derivative approxi-
mated by a 2nd order nite dierence approximation. The terms dropped in the Tay-
lor's series approximation of the cosine function provides higher accuracy in waveeld
extrapolation using the nite dierence representation of the time derivative. In fact,
for homogeneous media, using the cosine provides exact extrapolation regardless of

t.
In the pseudo spectral approach, space derivatives are computed in the Fourier
domain. This amounts to a nite dierence approximation with order accuracy given
by the number of samples in each dimension. Thus, it results in extrapolations free
of dispersions considering that we have enough samples to properly represent the
wavelength. As a result, equation 15 has the form:
P (t +
t; x)  2P (t; x) P (t
t; x) + v 2 (x)(
t)2 IF F T fjkj F T T fP (t; x)gg: (16)
Unlike the spectral method, we can use Fast Fourier transforms (FFT) and the ex-
trapolation is generally dispersion free granted they are not aliased. However, the
accuracy of the extrapolation now depends on the time step even in homogeneous
media. The accuracy can be evaluated by comparing the exact phase needed to shift
the waveeld, given by the spectral implementation, and one corresponding to the
Taylor's series expansion. We can also evaluate the amplitude to ensure that the
Alkhalifah 5 Spectral methods

pseudo spectral approximate approach does not alter the amplitude by much, have
an amplitude close to one. Using more terms in the expansion of the cosine function
in equation A-4 yields higher accuracy, but requires more FFT's. Accuracy tests are
provided later.
To use the highly accurate spectral method in inhomogeneous media, we have
to nd ways to reduce the cost of the 3-dimensional integral calculation granted the
mixed domain nature of the phase operator. This is done by nding ways to sepa-
rate the space dependency from the wavenumber dependency in the phase operator,
which will allow us to use FFTs. Pseudo spectral methods are one type of approxi-
mation. Next we investigate other approximations that retain the spectral approach
categorization.

Phase shift plus interpolation

Since the cost of evaluating equation 11 for every time step is expensive for inhomo-
geneous media, we can extrapolate many homogeneous media versions and then use
the velocity model to interpolate between these solutions. The number of forward
and inverse FFT's depend on the number of homogeneous media extrapolations. Et-
gen and Brandsberg-Dahl (2009) used two velocities, specically the maximum and
minimum velocities in the model. Thus, equation 11 is evaluated twice per time step
to obtain P1(t +
t; x) and P2(t +
t; x) corresponding to the minimum (vmin) and
maximum velocity (vmax), respectively. We then use the velocity model, v(x), with
the following formula based on linear interpolation
v (x)
v
P (t +
t; x)  P (t +
t; x) max
1 vmax vmin
+ P (t +
t; x) v(x) vmin ; ; (17)
2 vmax vmin
granted that the medium is not homogenous.
A split step implementation is also possible at a cost of a single 3D forward and
inverse Fourier transform per time step.

Variable separation methods

The spectral approach given by form 11 requires expensive integral evaluations every
time step. In 3D, the cost is of an of order O(n6), which is prohibitive. Note for a
constant velocity, the integrals become classical inverse Fourier transforms in which
we can use Fast divide and conquer methods, like those we learned earlier and the
cost reduces to O(n3(log n)3), which is much more forgiven.
The key idea of the variable seperation is decomposing the wave extrapolation
matrix
W (x; k) = ei [(x;k;
t) k
x] (18)
Alkhalifah 6 Spectral methods

for a xed
t into a separated representation

XM X N
W (x; k)  W (x; km )amn W (xn ; k): (19)
m=1 n=1
Representation (19) speeds up the computation of P (x; t +
t) since
Z
P (x; t +
t) = eixk W (x; k)Pb (k; t)dk
M
X N
X Z !
 W (x; km ) amn (20)
eixk W (xn ; k)Pb (k; t)dk :
m=1 n=1
The evaluation of the last formula is eectively equivalent to applying N inverse Fast
Fourier Transforms. Physically, a separable lowrank approximation amounts to se-
lecting a set of N representative spatial locations and M representative wavenumbers.
In order to discuss the construction of approximation (19), let us view it as a
matrix decomposition problem
W  W1 A W2 (21)
where W is the Nx  Nx matrix with entries W (x; k), W1 is the submatrix of W that
consists of the columns associated with fkmg, W2 is the submatrix that consists of
the rows associated with fxng, and A = famng. In practice, we nd that the matrix
W has a low rank separated representation provided that
t is suciently small,
which, in the case of smooth models, can be partially explained by the separation of
terms in the Taylor series 9. There are many ways to implement such a decomposi-
tion. Probably the most obvious is the singular value decomposition (SVD). However,
computing SVD for this potentially large matrix is expensive. Considering we only
need the eigenvalues and eigenvectors corresponding to the few largest eigenvalues,
maybe even, the largest as a good approximation, there are ways to speed up the com-
putation. However, we will discuss another approach that is new, fast, and extremely
accurate referred to as the low rank method.

The low rank approach

Let " be a prescribed accuracy of this separated representation, and r" be the numer-
ical rank of W. The construction of the separated representation in equation (21)
follows the method of Engquist and Ying (2007, 2009) and is detailed in the appendix.
The main observation is that the columns of W1 and the rows of W2 should span the
column space and row space of W, respectively, as well as possible. The algorithm
for computing takes the following steps:
1. Pick a uniformly random set S of 
r" columns of W where  is chosen to be 3
or 4 in practice. Perform the pivoted QR factorization to (W(:; S )). The rst
r" pivoted columns correspond to r" rows of the matrix W(:; S ). Dene W1
to be the submatrix of W that consists of these rows and set x1; : : : ; xN with
n = r" to be the corresponding x values of these rows.
Alkhalifah 7 Spectral methods

2. Pick a uniformly random set T of 
r" rows of W and perform the pivoted QR
factorization to W(T; :). Dene W2 to be the submatrix of W that consists of
these columns and set k1; : : : ; kM with m = r" to be the corresponding k values
of these columns.
3. Set the middle matrix A = Wy(xn; km)1nN;1mM where y stands for the
pseudoinverse.
4. Combining the result of the previous three steps gives the required separated
representation W  W1 A W2.
The algorithm does not require, at any step, access to the full matrix W, only to its
selected rows and columns. Once the decomposition is complete, it can be used at
every time step during the wave extrapolation process. In multiple-core implementa-
tions, the matrix operations in equation (19) are easy to parallelize.

A RESIDUAL SPECTRAL APPROACH

Now we look at other opportunities in gaining eciency in the spectral extrapolation

approach. The wave equation in the Fourier space-wavenumber domain reduces to a
rst-order ordinary dierential operator in one-way time. An approximate solution
equips us with the wave extrapolation formula, valid for small time step,
t, as follows
(Wards et al., 2008)
Z
P (t +
t; x)  Pb (t; k) ei (x;k)
t+i k
x dk; (22)
where Pb (t; k) is the waveeld in the wavenumber domain given by k = (kx; ky ; kz )
at time t, andp thus, x = (x; y; z) is a vector representing the space coordinates in
3D, and i = 1. In the geometrical (high-frequency) approximation, the function
(x; k) appearing in equation (22), for isotropic media, is given by
q
 = v jkj = v (x) kx2 + ky2 + kz2 : (23)
The linear nature of the wave equation allows us to propagate the full waveeld
or part of it to allow for the following representation
Z Z
P (t+
t; x) = P0 (t+
t; x)+
P (t+
t; x)  P0 (t; k) e i  ( x ;k )
t+i k
x dk+
P (t; k) ei (x;k)
t+i k
x dk;
(24)
where
P (t; k) is part of the waveeld that satises the wave equation for the same
velocity model but a source based on the perturbation. I refer to this part as the
residual waveeld.
A perturbation in the velocity model results in a perturbation in the extrapolation
operator and more precisely in the phase,
. Thus,
Z
P (t +
t; x)  P (t; k) ei 
t ei

t ei k
x dk;
0
(25)
Alkhalifah 8 Spectral methods

where 0 is the phase corresponding to the background medium (velocity, v0). Taking

 to be small, we can expand the exponential in equation 25 that includes
 using
any one of the many series or rational approximations, some will be discussed in the
next section, but most admit the following form:
ei

t = 1 + F (

t); (26)
where F includes the expansion terms as a function

t, using any approach in-
cluding Taylor's series (Etgen, 1989). Inserting equation 26 into equation 25 yields,
Z Z
P (t +
t; x)  P (t; k) ei 0
t+i k
x dk + P (t; k) F (

t) ei 0
t+i k
x dk: (27)
Combining formulas 24 and 27, I obtain a standalone residual waveeld extrapolation:
Z Z

P (t +
t; x) 
P (t; k) ei 
t+i k
x dk +
0
P (t; k) F (

t) ei 0
t+i k
x dk; (28)
where P = P0 +
P . Note that the second term on the right hand side need to
be computed only when
 6= 0. On the other hand, the rst term propagates the
residual waveeld using the background, possibly homogenous, model represented by
0 . This is equivalent to implementing a Born-like approximation at every time step.
Though Born has its limitations in the velocity perturbation, considering the small
time step nature of this implementation the impact of these limitations are heavily
mitigated to allow for a Huygen's principle type wave propagation.
If the background waveeld is obtained for a homogeneous medium, which can
be done analytically, the residual waveeld will approximately address the inhomo-
geneity. Up to this point there is no advantage to extrapolating the residual operator
instead of the full one, other than its potential usage for velocity estimation.

THE EXPANSION

To derive the residual exponential operator using a form that allows us to separate the
background contribution from the residual one, we utilize Taylor's series expansion (to
the second order) of the residual exponential extrapolation operator (Etgen, 1989),
as follows
F = ei
t
 1 = i
t


t2
2 1 i
t3
3+
t4
4 + 1 i
t5
5
t6
6 +O 
7 :
2 6 24 120 720
(29)
Truncating F to only the rst term amounts to a split-step type implementation.
Using this polynomial based expansion allows us to readily separate the space depen-
dence part (residual velocity) from the wavenumber part (the Laplace operator), in
a form commonly referred to as the pseudo-spectral implementation (Reshef et al.,
1988b). The accuracy of the residual extrapolation relies on the order of the expan-
sion, the amount of velocity residual and the time step of the extrapolation.
Alkhalifah 9 Spectral methods

To improve the accuracy of the Taylor's series expansion I use Shanks transform
to predict the higher-order behavior of the series by introducing a denominator, thus
requiring fewer terms. The rst-order Shanks transform (Bender and Orszag, 1978)
admits the following form
i

F = ei
t
 1 = (30)
1 i
t2
 :
This form, unlike Taylor's series, does not allow us to factor out the velocity requiring
a full spectral implementation. We will later address this issue through additional
approximations. Pade approximations are also possible (Zhang et al., 2004) for the
spectral implementations, but they do not admit a series form required by the residual
formulation given by equation 26.

ACCURACY TESTS

Considering an isotropic homogeneous model with a velocity of 3 km/s, I analyze

the accuracy of computing the phase of the extrapolation operator using the various
approximations of the residual operator considering a background velocity of v =
2km=s and as a result a velocity perturbation of v = 1km=s. Despite the relatively
large perturbation, I consider a time step of 0.01 s (of course this time step will have
to be reduced considerably for extrapolation in inhomogeneous media). Figure 1
demonstrates the accuracy of the approximations given in the previous section, and
specically the accuracy of the Shanks transform representation (dashed black curves)
given by equation 30 as a function of radial wavenumber (a) and a function of velocity
perturbation (b). Errors increase with an increase in the wavenumber or an increase in
the velocity perturbation. Nevertheless, the increase in errors as a function of velocity
perturbation is exponential, which suggests that a residual implementation with small
steps in velocity perturbation is worth exploring. Clearly all approximations are
extremely accurate for small velocity perturbations or equivalently small time steps,
with errors increasing exponentially with an increase in these two variables. Since a
larger time step speeds up the extrapolation to the limit that velocity variations allow
us, we focus on the velocity perturbation. The sixth-order Taylor's series expansion
is extremely accurate but it costs a lot. With the smaller time step needed to handle
complex velocity models the accuracy dierence between the sixth order and the rst
order becomes much smaller. Shanks transform provided high accuracy, higher than
the second order with fewer terms.
However, Shanks transform does not allow for straightforward separation of the
velocity for a pseudo-spectral implementation. A simple remedy is provided by replac-
ing
 in the denominator of equation 30 by a constant value given by the average
velocity perturbation. Considering we are dealing with controllable perturbations, the
deviation introduced by this approximation can be small. Figure 2 demonstrates that
changes to the perturbation velocity in the denominator, negatively (a) or positively
(b), still yields higher accuracy than the second-order approximation. A constant
Alkhalifah 10 Spectral methods

a b
0.10 0.10

0.05 0.05
Φ ErrorH%L

Φ ErrorH%L
0.00 0.00

-0.05 -0.05

-0.10 -0.10
0 5 10 15 20 -1.0 -0.5 0.0 0.5 1.0
-1
kr Ikm M DΥHkmsL

Figure 1: Percent errors in the phase of the extrapolator operator given by the rst-
order (solid black curve) and second-order (solid grey curve) expansions, rst-order
Shanks transform 30 (dashed black curve), and sixth-order Taylor's series expansion
(dashed grey curve, almost zero), as a function of wavenumber for a perturbation

v = 1km=s (a) and perturbation velocity for kr = 10km 1 (b). The model has a
velocity of 2 km/s and a time step of 0.01 s.

average velocity for the denominator reduces the number of required inverse Fourier
transforms for the perturbation part to one instead of 2 for the second-order expan-
sion.
We can also evaluate the amplitude of the approximated phase operator, which
should be one. Figure 3 shows such amplitude for the second-order approximation.
We have to use a time step of 0.02 s (Figure 3c), which is large, to see some errors
especially in the high wavenumber components. Shanks transform provides exact
amplitudes.

SMALLER CONSTANT-VELOCITY STEPS

The residual implementation can be used to apply as many residuals as we want.

Wards et al. (2008) investigated formulations for a single velocity perturbation in a full
implementation (not a residual one). To demonstrate the advantages of considering
small successive velocity perturbations, we divide the perturbation in velocity to n
smaller perturbations each corresponding to
vi =
nv . Considering the rst term of
a Taylor's series expansion
n
Y
ei
t = ei
t0 ei
t
i
i=1
Yn
= ei
t0 (1 + i
t
i)
i=1
Alkhalifah 11 Spectral methods

a b
0.10 0.10
0.08 0.08
Φ ErrorH%L

Φ ErrorH%L
0.06 0.06
0.04 0.04
0.02 0.02
0.00 0.00
0 5 10 15 20 0 5 10 15 20
-1 -1
kr Ikm M kr Ikm M

Figure 2: Percent errors in the phase of the extrapolator operator given by the
Shanks transform expansion 30 (solid black curve), second-order Taylor's series ex-
pansion(solid grey curve), Shanks transform with the perturbation velocity o by 20%
(dashed black curve), and the perturbation velocity o by 50% (dashed grey curve),
in the negative direction (a) and the positive direction (b). The model has a velocity
of 2 km/s,
v = 1km=s and a time step of 0.01 s.

a b c
1.003 1.003 1.003

1.002 1.002 1.002

Amplitude

Amplitude

Amplitude

1.001 1.001 1.001

1.000 1.000 1.000

0.999 0.999 0.999

0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
kr Ikm-1 M kr Ikm-1 M kr Ikm-1 M

Figure 3: Amplitude of the exact phase operator (solid curve), equal 1, compared
to that extracted using the second-order Taylor's series expansion in the residual
velocity (dashed curve)., for (a)
t = 0:004s, (b)
t = 0:01s, and (c)
t = 0:02s.
The velocity is equal to 2 km/s and
v=1 km/s.
Alkhalifah 12 Spectral methods

0 ! 0 1
n
X X
= ei
t0 @1 + i
t
i +
t2 @
i
j A
i=1 (i6=j )
1
X
i
t3
i
j
k + R (
t; n)A ; (31)
(i6=j 6=k)
where
i =
vijkj, n is the number of velocity discretization terms (n > 2), and
Q
stands for the product symbol. The function R contains the other terms of the
expansion. For n = 3, R = 0, and for n > 3, R contains terms up to the order

tn. For n =1 2, only the rst two terms of the expansion remain. In this case,

v1 =
v2 = 2
v, and thus
F  i
t

t2
2 : (32)
Compared to the Taylor's series expansion 30, we are missing the division over two in
the second term. A similar result is obtainable by dividing the time step to articially
smaller ones, but the velocity implementation allows for more options as velocity may
vary with position.
We now test the accuracy of this new implementation. To reduce the number of
inverse Fourier transforms required I will investigate the cases where n=2 and n=4.
For n = 4, we also truncate the product expansion back to two terms stopping at

t2, and thus only consider the rst three terms in brackets in equation 31, which is
equivalent to a second order approximation. Figure 4a shows the percent error in the
phase operator as a function of wavenumber. Among the second-order equivalent (in
cost) implementations, the truncated four-step (n = 4) implementation (dashed grey
curve) has the highest accuracy. This is appreciated even more in Figure 4b where
the errors in the operators are given as a function of the velocity perturbation.

WAVEFIELD EXTRAPOLATION

Using the low rank approach of Fomel et al. (2010) to deal with the mixed space-
wavenumber domain nature of the phase operator function in equation 22, we extrap-
olate waves in inhomogeneous media. The rank of the separable operators (space and
wavenumber dependent) control the number of inverse Fourier transforms required
in each extrapolation step, and thus, the cost. This is typically dependent on the
complexity of the velocity model and the size of time step used in the extrapolation.
We apply the residual extrapolator on the BP model (Billette and Brandsberg-
Dahl, 2004) using a Ricker-wavelet at a point source with a maximum frequency of
50 Hz. The horizontal grid size
x is 37.5 m, the vertical grid size
z is 12.5 m and
the time step is 1 ms. Figure 5 shows the part of the model with a salt body.
A snap shot of the waveeld at time 3.2 s extracted from waveeld extrapolation
using the low rank method applied to the exact operator given by equation 22 is shown
in Figure 6(a). We compare this reference waveeld with that obtained using the
Alkhalifah 13 Spectral methods

a b
0.10 0.10

0.05 0.05
Φ ErrorH%L

Φ ErrorH%L
0.00 0.00

-0.05 -0.05

-0.10 -0.10
0 5 10 15 20 -1.0 -0.5 0.0 0.5 1.0
-1
kr Ikm M DΥHkmsL

Figure 4: Percent errors in the phase of the extrapolator operator given by the
second-order expansion (solid black curve) and two-step velocity perturbation, which
is equivalent to second order (solid grey curve), four-step velocity perturbation, which
is equivalent to fourth order (dashed black curve), and the four step truncated to
second order (dashed grey curve), as a function of wavenumber for a perturbation

v = 1km=s (a) and as a function of the perturbation velocity for kr = 10km 1 (b).
The background model has a velocity of 2 km/s and an extrapolation time step of
0.01 s.

residual extrapolator based on a linear approximation of the Taylor's series expansion

in equation 30 (using only the rst term on the right hand side); a snap shot at
3.2 s shown in Figure 6(b). The dierence plotted at the same scale is shown in
Figure 6(c). The dierence between the reference exact and using Shanks transform
instead (which requires the same number of pseudo-spectral inverse Fourier transform
as the linear) plotted again at the same scale is shown in Figure 6(d). Clearly, the
dierence associated with the Shanks transform implementation, with the same cost,
is lower.
Another comparison between the dierence of the exact operator and the linear
approximation (shown in Figure 7(a)), and that of the exact operator and the modi-
ed Shanks transform that allows for a residual implementation (Figure 7(b)), both
plotted at a 10x gain compared to Figures 6(a)-6(d). The errors in the modied
Shanks transform are clearly smaller. Considering that both extrapolations cost the
same, the added accuracy with the modied Shanks transform implementation is
helpful for many applications.

SPECTRAL METHODS COST

Though the emphasis here is on accuracy, a residual implementation has potential cost
advantages. Considering a background time extrapolation operator, 0, for homoge-
neous or smoothly inhomogeneous media, the extrapolation of such operators tend
Alkhalifah 14 Spectral methods

Figure 5: Portion of BP 2004 synthetic velocity model.

to be cheap, as approximations in the spectral implementation requires less inverse

Fourier transforms per time step (or, in other words, a lower rank representation,
especially for a relatively large time step). Thus, using the residual implementation
given by equation 28, we need to evaluate the second term, with the extrapolation
operator for the more complex medium, only when residuals in the medium exists
(
v 6= 0). A single combined extrapolation operator will not recognize such simplic-
ity in the model in certain areas (like within a salt body). This opens the door for
many smart velocity separation based operators that can help in optimally improving
the speed of the extrapolation.

REFERENCES

Alkhalifah, T., 1998, Acoustic approximations for processing in transversely isotropic

media: Geophysics, 63, 623{631.
Baysal, E., D. D. Koslo, and J. W. C. Sherwood, 1983, Reverse time migration:
Geophysics, 48, 1514{1524.
Bender, C. M., and S. A. Orszag, 1978, Advanced mathematical methods for scientists
and engineers: McGraw-Hill.
Billette, F. J., and S. Brandsberg-Dahl, 2004, The 2004 BP velocity benchmark: 67th
Annual EAGE Meeting, EAGE, Expanded Abstracts, B305.
Alkhalifah 15 Spectral methods

(a) (b)

(c) (d)

Figure 6: Snap shots of the waveeld at 3.2 s using the conventional extrapolation
operator, equation 22 (a), using the linear approximation (b), the dierence (c), and
the dierence between the exact operator and the Shanks transform approximation
(d) for the BP model shown in Figure 5, with a source near the surface.
Alkhalifah 16 Spectral methods

(a)

(b)

Figure 7: Snap shots of the dierence in the waveeld at 3.2 s (a) between the exact
operator and linear approximation, and (b) and the Shanks transform approximation
with the constant average velocity used in the denominator to allow for the residual
implementation. Both plots are at the same scale, 10 times that in Figures 6(a)-6(d).
The cost of implementing both operators is the same.
Alkhalifah 17 Spectral methods

Chu, C., and P. Stoa, 2011, Application of normalized pseudo-laplacian to elastic

wave modeling on staggered grids: Geophysics, 76, T113{T121.
Etgen, J., 1989, Accurate wave equation modeling: SEP-60, 131{148.
Etgen, J., and S. Brandsberg-Dahl, 2009, The pseudo-analytical method: application
of pseudo-Laplacians to acoustic and acoustic anisotropic wave propagation: 79nd
Annual International Meeting, SEG, Expanded Abstracts, 2552{2556.
Fletcher, R. P., X. Du, and P. J. Fowler, 2009, Reverse time migration in tilted
transversely isotropic (TTI) media: Geophysics, 74, WCA179{WCA187.
Fomel, S., L. Ying, and X. Song, 2010, Seismic wave extrapolation using lowrank
symbol approximation: SEG Technical Program Expanded Abstracts, 29, 3092{
3096.
Fowler, P. J., X. Du, and R. P. Fletcher, 2010, Coupled equations for reverse time
migration in transversely isotropic media: Geophysics, 75, S11{S22.
Geller, R., and N. Takeuchi, 1998, Optimally accurate second-order time- domain
nite dierence scheme for the elastic equation of motion: one- dimensional case:
Geophys. J. Int., 135, 4862.
Kristek, J., P. Moczo, and M. Galis, 2010, Stable discontinuous staggered grid in the
nite-dierence modelling of seismic motion: Geophys. J. Int., 183, 14011407.
Levin, S. A., 1984, Principle of reverse time migration: Geophysics, 49, 581{583.
Liu, Y., and M. K. Sen, 2009, A new time-space domain high-order nite-dierence
method for the acoustic wave equation: Journal of computational Physics, 228,
8779{8806.
McMechan, G. A., 1983, Migration by extrapolation of time-dependent boundary
values: Geophys. Prosp., 31, 413{420.
Pestana, R. C., and P. L. Stoa, 2010, Time evolution of the wave equation using
rapid expansion method: Geophysics, 75, T121{T131.
Reshef, M., D. Koslo, M. Edwards, and C. Hsiung, 1988a, Three-dimensional acous-
tic modeling by the fourier method: Geophysics, 53, 1175{1183.
||{, 1988b, Three-dimensional acoustic modeling by the Fourier method: Geo-
physics, 53, 1175{1183.
Soubaras, R., and Y. Zhang, 2008, Two-step explicit marching method for reverse
time migration: 78th Ann. Internat. Mtg., Soc. Expl. Geophys., 2272{2276.
Symes, W. W., 2007, Reverse time migration with optimal checkpointing: Geophysics,
72, SM213{SM221.
Tal-Ezer, H., D. Koslo, and Z. Koren, 1987, An accurate scheme for seismic forward
modeling: Geophysical Prosecting, 35, 479{490.
Wards, B. D., G. F. Margrave, and M. P. Lamoureux, 2008, Phaseshift timestepping
for reversetime migration: SEG, Expanded Abstracts, 27, 2262{2266.
Whitmore, N. D., 1983, Iterative depth migration by backward time propagation:
53rd Ann. Internat. Mtg., Soc. Expl. Geophys., 382{385.
Wu, W., L. R. Lines, and H. Lu, 1996, Analysis of high-order, nite-dierence schemes
in 3-D reverse-time migration: Geophysics, 61, 845{856.
Yoon, K., K. J. Marfurt, and W. Starr, 2004, Challenges in reverse-time migration:
74th Ann. Internat. Mtg., Soc. Expl. Geophys., 1057{1061.
Zhang, H., G. Zhang, and Y. Zhang, 2009, Removing S-wave noise in TTI reverse
Alkhalifah 18 Spectral methods

time migration: 79th Ann. Internat. Mtg., Soc. Expl. Geophys., 2849{2853.
Zhang, L., J. W. Rector, G. M. Hoversten, and S. Fomel, 2004, Splitstep complex
padfourier depth migration: SEG, Expanded Abstracts, 23, 989{992.
Zhang, Y., and G. Zhang, 2009, One-step extrapolation method for reverse time
migration: Geophysics, 74, A29{A33.

EXPANSIONS WITH RESPECT TO THE COS

OPERATOR

A second-order time extrapolation of the waveeld is performed using

P (t +
t; x) + P (t
t; x)
Z
 2 Pb (t; k) cos [(x; k)
t] ei k
x dk : (A-1)
Considering perturbations in the velocity, the phase operator can be written in a
separable form as follows:
cos (0 +

t) = cos (0
t) cos(

t) sin (0
t) sin(

t): (A-2)
Consequently, for small
 and small
t, we obtain
cos (0
t +

t) = cos (0
t) (1 + F ) sin (0
t) E; (A-3)
where, using Taylor's series expansion,
F=

t2
2 +
t4
4
t6
6 + O 
7
2 24 720
E =
t


t3
3
t5
5
+ +

O
7 :

(A-4)
6 120
Again, considering P = P0 +
P , we obtain
Z

P (t +
t; x) +
P (t
t; x)  2
Pb (t; k) (cos (0
t) (1 + F ) sin (0
t) E ) ei k
x dk
Z
+2 Pb0(t; k) (cos (0
t) F sin (0
t) E ) ei k
x dk
Z Z  
 2
( ) cos (0
t)
Pb t; k +2
ei k
x dk Pb0 (t; k) +
Pb (t; k)
(cos (0
t) F sin (0
t) E ) ei k
x dk;
(A-5)
which describes the extrapolation of the residual eld, and it includes a term that
includes an extrapolation of the residual eld using the background operator and
another containing the interaction of the residual operator with the full waveeld.