Вы находитесь на странице: 1из 18

GEOPHYSICS, VOL. 66, NO. 5 (SEPTEMBER-OCTOBER 2001); P. 1551–1568, 23 FIGS.

Modeling and imaging with the scalar generalized-screen


algorithms in isotropic media

Jérôme H. Le Rousseau and Maarten V. de Hoop∗

ionosphere (Buckley, 1975; Bramley, 1977; Knepp, 1983), prop-


ABSTRACT agation of acoustic waves in the ocean (Flatté et al., 1979;
Thomson and Chapman, 1983), and propagation of seismic
The phase-screen and the split-step Fourier methods,
waves in the earth (renamed as the split-step Fourier method;
which allow modeling and migration in laterally hetero-
Stoffa et al., 1990). More recently, the phase-screen method
geneous media, are generalized here so as to increase
has been extended to elastic waves (Fisk and McCartor, 1991;
their accuracies for media with large and rapid lateral
Fisk et al., 1992; Wu, 1994).
variations. The medium is defined in terms of a back-
The original phase-screen method was designed for multiple
ground medium and a perturbation. Such a contrast
downward scattering of waves, the downward direction being
formulation induces a series expansion of the vertical
the preferred direction of propagation. It included phenomena
slowness in which we recognize the first term as the
such as focusing and defocusing associated with multipathing.
split-step Fourier approximation and the addition of
The applicability of the phase-screen method generally
higher-order terms of the expansion increases the ac-
requires that the screen interval satisfies the following crite-
curacy. Employing this expansion in the one-way scalar
ria: small medium variations (weak scattering), transversely
propagator yields the scalar one-way generalized-screen
smooth medium variations (narrow-angle scattering), and even
propagator. We also introduce a generalized-screen rep-
smoother variations in the preferred direction (negligible
resentation of the reflection operator. The interaction
backscattering). With the generalized-screen (GS) approach,
between the upgoing and downgoing fields is taken into
we access the accuracy of the phase-screen method and gener-
account by a Bremmer series. These results are then
alize it to larger contrast, wider angle, and backscattering.
cast into numerical algorithms. We analyze the accuracy
In realistic geological models, heterogeneity in medium
of the generalized-screen method in complex structures
properties is such that the phenomenon of multiple scattering
using synthetic models that exhibit significant multi-
is significant. We distinguish two classes of multiple scatter-
pathing: the IFP 2-D Marmousi model and the SEG-
ing: one in which the multiples are identified with respect to
EAGE 3-D salt model. As compared with the split-step
the projection of their propagation paths onto the preferred
Fourier method, in the presence of lateral medium vari-
or vertical direction (depth), and one where the multiples are
ations, the generalized-screen methods exhibit an in-
identified with respect to the projection of their propagation
creased accuracy at wider angles of propagation and
paths onto the transverse or horizontal plane. In the asymp-
scattering. As a result, in the process of migration, we
totic framework of wavefront analysis, paths are rays. The first
can choose a member of the family of our generalized-
class of multiple scattering is associated with “turning rays” and
screen algorithms in accordance with the complexity of
“internal multiples” as well as “surface multiples”; the second,
the medium (velocity model).
possibly combined with the first class of multiple scattering, is
associated with “multipathing.”
INTRODUCTION Seismic imaging and inversion are now commonly applied to
regions where geologic complexities are present. An important
The concept of screen approximations to the propagation part of the seismic energy is not contained in the “first” arrival.
of waves has been around for many years. The phase-screen An accurate prediction of multipathing and “second” arrivals
approximation has been applied to light transmission through is the key to a better processing. A ray-theoretic treatment of
the atmosphere (Ratcliffe, 1956; Mercier, 1962; Filice, 1984; the multipathing is not straightforward and is algorithmically
Martin and Flatté, 1988), propagation of light in optical fibers rather involved (De Hoop and Brandsberg-Dahl, 2000). Wave
(Feit and Fleck, 1978), propagation of radio signals through the extrapolation methods are able to predict multipathing (second

Manuscript received by the Editor September 29, 1999; revised manuscript received January 22, 2001.

Colorado School of Mines, Center for Wave Phenomena, Golden, Colorado 80401-1887. E-mail: jlerous@dix.mines.edu; mdehoop@dix.mines.edu.
°c 2001 Society of Exploration Geophysicists. All rights reserved.

1551
1552 Le Rousseau and de Hoop

class of multiple scattering), with no need to follow the forma- We analyze the accuracy of the GS method in complex
tion of caustics explicitly. At the same time, with 3-D surveys structures using synthetic models that exhibit significant mul-
becoming standard practice, fast 3-D algorithms are in demand. tipathing: the IFP 2-D Marmousi model and the SEG-EAGE
On the one hand, the full 3-D extension of finite-difference 3-D salt model. These two models represent two fundamen-
methods is costly and restrict their application. On the other tally different geological situations. In the Marmousi model,
hand, methods such as the phase-screen (Ratcliffe, 1956) and complexity arises from faulting and tectonic deformation in
the closely related split-step Fourier (Stoffa et al., 1990) meth- a sedimentary region. In the SEG-EAGE 3-D salt model, it
ods yield a faster 3-D algorithm. They are, however, limited arises from the intrusion of a salt body, the wave speed of
in their capacity to predict large-angle propagation where sig- which is significantly higher than in the surrounding forma-
nificant lateral heterogeneities are present. Because of their tions. These models commonly yield poor imaging below these
attractive properties (3-D, multipathing), we propose here to complex structures. With the help of the GS propagator, which
generalize this family of algorithms, enhancing their accuracy. we prove to be accurate in these situations, we shall illustrate
Since we preserve the algorithmic structure of the one of the that the origin of this problem is possibly associated with mul-
classical phase-screen propagator, we denote our approxima- tipathing.
tions as generalized screens. We first present the GS representation of the thin-slab
In addition, approach accounts for the first class of multiple propagator; in Appendix C, we develop the GS expansion
scattering through use of the generalized Bremmer series (De of reflection operator. These results are then cast into nu-
Hoop, 1996). merical algorithms, and our accuracy analysis is carried out
A scattering theory that follows the ray picture but accounts primarily by modeling. We analyze the migration operator
for full-wave behavior has been developed by De Hoop (1996). before stacking that is conventionally performed in the pro-
It is based on an extension of the Bremmer coupling series to cess of imaging. We focus on multipathing and “second”-
multidimensionally varying media. Bremmer’s method decom- arrival energy. We illustate the modeling capacity of the
poses the wavefield into a recursion of one-way propagation GS method in two and three dimensions. For complete-
operators each using the previous wavefield as a source. Thus, ness, we show some prestack depth migration results in two
the method first generates a wavefield dominated by down- dimensions.
ward propagation, then generates a “first” upward propagat- Throughout this paper, we assume that the medium is
ing wavefield, then a “second” downward propagating field, isotropic. In a companion paper (Le Rousseau and De Hoop,
etc. In this manner, multiple reflections (class one multiples) 2001), we extend the method to media that are transversely
are accounted for in a controlled manner. The Bremmer series isotropic with vertical symmetry axis.
in acoustic media has already found application in Van Stralen
et al. (1998), where the one-way wave operator is approximated THE SCALAR GENERALIZED-SCREEN PROPAGATOR
with an accurate optimal finite-difference algorithm based on
a rational expansion as opposed to a GS expansion. The scalar one-way Green’s function
The propagator that generates the Bremmer series can be Selecting the direction of preference along the x 3 -axis (or
represented by a Hamiltonian path integral (De Witte-Morette “vertical” axis) and denoting the remaining (“transverse” or
et al., 1979; Fishman and McCoy, 1984a, b; De Hoop, 1996) “horizontal”) coordinates by xµ , µ = 1, 2, the one-way acous-
that accounts for not only the energy traveling along the ray tic Green’s function G (±) is defined by the equation
but also for the transport along nonstationary paths. These path
integrals contain all possible multipathing. In the path integral, [∂3 ∓ iω0(xµ , x3 ; Dν )]G (±) (xµ , x3 ; xυ0 , x30 )
“time” is identified with depth, and “momenta” are identified
with the horizontal wave slownesses which, in the ray-theoretic = δ(xµ − xµ0 ) δ(x3 − x30 ) (1)
limit, coincide with the horizontal components of the gradient
of travel time. The (square-root) Hamiltonian, appearing in in the frequency-space domain. The one-way Green’s functions
the phase of the path integral, is identified with vertical wave propagate waves in an arbitrary inhomogeneous medium, ig-
slowness which, in the ray-theoretic limit, coincides with the noring class one multiples, taking class two multiples into ac-
vertical component of the gradient of travel time (De Hoop, count. The left-hand side of equation (1), ∂3 ∓ iω0, is the one-
1996). way wave operator. The choice of sign discriminates upgoing
The problem with the path integrals is the computational (G (−) ) from downgoing (G (+) ) Green’s functions. For uniform
complexity of their numerical evaluation. De Hoop et al. (2000) density ρ, the operator 0 satisfies the characteristic equation,
developed a method (GS) that dramatically reduces the com-
putational complexity of such evaluation, at the cost of ap- 0 2 = A, A = c−2 − Dν Dν , (2)
proximating the acoustics (the shape of wavefronts). The re-
sult is an algorithm that, for each propagation step, is built where c is the medium wave speed and
from the sequence: forward Fourier transform, multiplication, 1
inverse Fourier transform, multiplication—where the trans- Dν ≡ ∂ν , (3)

form is in the horizontal directions and may be windowed.
We have designed a hierarchy of increasingly accurate approx- has the interpretation of horizontal slowness operator; 0 =
imations. Underlying these approximations is an expansion of 0(xµ , x3 ; Dν ) is the vertical slowness operator. In equation (2)
the medium wave-speed model simultaneously into magnitude and throughout the paper, we use the summation convention
and smoothness of variation. for repeated indices.
Generalized Screens 1553

For variable density media, De Hoop (1996) derived that where the Heaviside function H generates the Dirac distribu-
tion in equation (1) (cf. Appendix A). For a sufficiently small
0 2 = −Dν Dν + κρ − ρ −1 (Dν ρ)Dν − ρ −1 (Dν Dν ρ) vertical step 1x 3 = x3 −x30 (thin slab) and a medium sufficiently
smooth, the Hamiltonian path-integral representation for the
+ ρ −2 (Dν ρ)(Dν ρ), (4) one-way thin-slab propagator reduces to (De Hoop et al., 2000)
Z
equal to the transverse Helmholtz operator in the acous-
g (±)
(xµ , x3 ; xυ0 , x30 ) '
(1/2π )2 dk1 dk2 exp[ikσ (xσ − xσ0 )]
tic pressure-normalization analog (De Hoop, 1996; De Hoop
et al., 2000); ρ is the volume density of mass (kg/m3 ), and κ Z
is the compressibility (Pa−1 ), with the wave speed defined as × exp[±iωγ (xµ , x̄3 ; ikν /ω)1x3 ] = (s/2π )2 dα1 dα2
c−2 = κρ.
× exp[−is ασ (xσ − xσ0 )] exp[∓sγ (xµ , x̄3 ; αν )1x3 ],
The thin-slab propagator.—In the present development, it
(8)
is advantageous to use the Laplace transform with respect to
time, t, and the Fourier transform with respect to the horizontal with
spatial coordinates, xν . We introduce the notation
x̄3 = x3 − 12 1x3 = x30 + 12 1x3 .
s = −iω, (5)
In the limit of a laterally homogeneous thin slab, γ will
1 1
αν = kν = − kν , (6) not depend on xµ , and the thin-slab propagator reduces to
iω s Gazdag’s phase-shift operator (Gazdag, 1978). The operator is
where ω and kν are the frequency and the horizontal wavenum- composed of a forward Fourier transform, a multiplication by
ber components (the inverse Laplace transform is anticipated a phase factor (the phase is proportional to the vertical slow-
to be evaluated along the Bromwich contour). In the Laplace ness), and an inverse Fourier transform. In the general case
(s) domain with Re(s) > 0, the operator A becomes strictly of equation (8), the thin-slab propagator has a similar struc-
elliptic, which enables us to consider any of its fractional pow- ture except that the phase factor is dependent upon the output
ers with the aid of pseudodifferential calculus (Treves, 1980). point, xµ . Every output point requires its own evaluation of
Here, we are interested in its square root, 0. Using the Fourier equation (8), which represents a considerable computational
transform with respect to the horizontal spatial coordinates, we effort. The GS approximation of the thin-slab propagator en-
can associate with 0 and A their “left symbols” (Treves, 1980; forces a simplification of this computational complexity while
Hörmander, 1985), γ and a, respectively; i.e., let φ be a test allowing laterally varying media.
function, then
Z Z High-frequency approximation.—Throughout this paper,
(0φ)(xµ , x3 ) = 2
(s/2π) dα1 dα2 dx10 dx20 γ (xµ , x3 ; αν ) we will use a “high-frequency” approximation. With it, the
pseudodifferential operator 0 reduces to its principal part 01 ,
× exp[−is(xσ − xσ0 )ασ ] φ(xυ0 , x3 ), and the associated left symbol, γ , reduces to its principal sym-
Z Z bol, γ1 . The principal symbol, γ1 , follows from taking the high-
(Aφ)(xµ , x3 ) = (s/2π) dα1 dα2 dx10 dx20 a(xµ , x3 ; αν )
2 frequency limit of equation (7) (cf. Appendix B) with solution
q
× exp[−is(xσ − xσ0 )ασ ] φ(xυ0 , x3 ), γ1 (xµ , x3 ; αν ) = [c(xµ , x3 )]−2 + αν αν . (9)
The triplet (iα1 , iα2 , γ1 ) represents the components of the gra-
where the symbol a is obtained, after equation (4), as dient of the travel time. Then, the scalar one-way thin-slab
propagator becomes
a = αν αν +κρ −ρ −1 (Dν ρ) iαν −ρ −1 (Dν Dν ρ)+ρ −2 (Dν ρ)(Dν ρ), Z
since iαν is the left symbol of Dν . For uniform density, a re-
g (±)
(xµ , x3 ; xυ0 , x30 ) ' (s/2π )2 dα1 dα2
duces to αν αν + κρ, consistent with equation (2). Equation (2)
× exp[−is ασ (xσ − xσ0 )] exp[∓sγ1 (xµ , x̄3 ; αν )1x3 ].
transforms into a characteristic equation for left symbols:
£ ¤ (10)
exp −i∂ασ0 Dxσ0 γ (xµ , x3 ; ασ0 )γ (xσ0 , x3 ; αν ) |(xµ0 ,αν0 )=(xµ ,αν )
Note that the wavefront set (Hörmander, 1983) of propaga-
= a(xµ , x3 ; αν ), (7) tor (10) is equal to the one of propagator (8). Singularities
of the wavefield propagate along the characteristics which are
as given by symbol calculus of pseudodifferential operators solely determined by the principal parts of the pseudodifferen-
[Treves, 1980; Hörmander, 1985; De Hoop, 1996, his equa- tial equation (1) (Hörmander, 1985, theorem 18.1.28). Using
tion (A12)]. the principal part only, we do not alter the geometrical aspects
With the vertical slowness left symbol, the one-way Green’s of the wave propagation (i.e., the wavefront set). Any depar-
function, G (±) , from equation (1) can be represented by a ture from the principal parts results in smoother contributions
Hamiltonian path-integral representation (Cohen-Tannoudji to the wavefield.
et al., 1977; Schulman, 1981; Fishman and McCoy, 1984b;
De Hoop et al., 2000). We define the scalar one-way propa- Generalized-screen principal-slowness surface
gator, g (±) , through
The contrast formulation.—For the subsequent analysis, we
G (±) = ±H (±[x3 − x30 ]) g (±) , employ a “contrast formulation” that allows us to take lateral
1554 Le Rousseau and de Hoop

heterogeneity into account in the thin-slab propagation. In the It is important to note that in each term of the expansion, the
slab [x30 , x3 ], we introduce a background medium with wave dependencies on xµ and on αν are separated. This property
speed c0 . The background medium is constant in the slab, but will induce the structure of the GS propagator. Justification of
may vary from one slab to another. We express this by letting expansion (16) is given in De Hoop et al. (2000), where the full
c0 = c0 (x3 ). A medium perturbation term u is then introduced symbol γ is analyzed; here, we have restricted ourselves to its
as principal part, γ1 . Where normal mode expansions of the one-
£ ¤−2 £ 0 ¤−2 way propagator apply, the separation, in fact, occurs naturally
u(xµ , x3 ) = c(xµ , x3 ) − c (x3 ) . (11)
(Fishman et al., 2000).
To avoid having an artificial branch point enter the propagating-
wave regime (see below), we impose the condition
Phase-screen and split-step Fourier approximations.—The
c0 (x3 ) ≤ c(xµ , ζ ) for ζ ∈ [x30 , x3 ]. (12) phase-screen approximation follows from expansion (16) by
setting n = 1 and approximating 1/γ 0 by its zero-order Taylor
We will expand the principal symbol of the vertical slowness
expansion in αν about 0 (vertical propagation):
left symbol, γ1 , into the perturbation u about the background
medium. γ1 (xµ , x3 ; αν ) ' γ 0 (x3 ; αν ) + 12 c0 u(xµ , x3 ). (19)

Generalized-screen expansion.—Assuming small vertical This leads to the approximation of the vertical slowness symbol
medium variation across the thin slab (i.e., if the thin slab is employed by Stoffa et al. (1990), since
sufficiently small), we set
q u = c−2 − (c0 )−2 ' 2(c0 )−1 (c−1 − (c0 )−1 ),
£ ¤−2
γ 0 (ζ ; αν ) = c0 (x3 ) + α ν αν which yields with equation (19),
£ ¤−1 £ 0 ¤−1
= γ 0 (x3 ; αν ) for ζ ∈ [x30 , x3 ], (13) γ1 (xµ , x3 ; αν ) ' γ 0 (x3 ; αν ) + c(xµ , x3 ) − c (x3 ) ,
the vertical slowness associated with the background medium. (20)
The principal symbol of the vertical slowness, γ1 , can then be
decomposed into a background component, γ 0 , and a pertur- as found in the split-step Fourier method.
bation, γ11 , i.e., Figure 1 illustrates the principal slowness√surfaces (γ1 as
a function of the horizontal slowness p = −αν αν ) of the
γ1 (xµ , x3 ; αν ) = γ 0 (x3 ; αν ) + γ11 (xµ , x3 ; αν ). (14) GS expansions for n = 1, 2 as given in equation (16), and
To find an explicit form for γ11 , we use relationship (9), which the split-step Fourier one. The split-step approximation ap-
can be rewritten as proaches the actual slowness surface by simply vertically
q shifting the background slowness surface. Adding higher-order
£ ¤−2
γ1 (xµ , x3 ; αν ) = c0 (x3 ) + u(xµ , x3 ) + αν αν terms in the GS expansion, the shape of the slowness surface
s is improved and, hence, the accuracy for wider-angle propaga-
u(xµ , x3 ) tion is increased. This observation indicates that for the exten-
= γ 0 (x3 ; αν ) 1 + £ ¤2 , (15) sion to anisotropic media, the split-step approximation does
γ 0 (x3 ; αν ) not have the degree of freedom to shape the slowness surface
appropriately (L. Thomsen, personal communication, 1998; Le
and expand the result in a Taylor series according to
Rousseau and De Hoop, 2001); the influence of the anisotropic
γ1 (xµ , x3 ; αν ) = γ 0 (x3 ; αν ) + γ11 (xµ , x3 ; αν ) parameters occurs only at nonvertical propagation.
The GS expansion as shown in equation (16) reveals the
X
n
[u(xµ , x3 )] j introduction of reciprocal powers of γ 0 and, hence, contains
= γ 0 (x3 ; αν ) + aj £ ¤2 j−1 + o[u(xµ , x3 )] ,
n

j=1 γ (x3 ; αν )
0

(16)
where
1 · 3 · · · (2 j − 3)
a j = (−1) j+1 . (17)
j! 2 j
In practice, we shall limit ourselves to the fourth-order
expansion,
1 u(xµ , x3 )
γ1 (xµ , x3 ; αν ) = γ 0 (x3 ; αν ) +
2 γ 0 (x3 ; αν )
1 [u(xµ , x3 )]2 1 [u(xµ , x3 )]3
− £ ¤3 + £ ¤ FIG. 1. Principal parts of the generalized-screen vertical slow-
8 γ 0 (x3 ; αν ) 16 γ 0 (x3 ; αν ) 5 ness: zero-order (GSP0), first-order (GSP1), and second-order
(GSP2) as a function of the horizontal slowness p. Also shown
5 [u(xµ , x3 )]4 is the principal part of the vertical slowness for the split-step
− £ ¤ + o[u(xµ , x3 )] .
4
(18) Fourier method. The principal part of the exact vertical slow-
128 γ 0 (x3 ; αν ) 7
ness is shown with the inner dashed curve.
Generalized Screens 1555

branch points at αν αν = −[c0 ]−2 , as illustrated in Figure 1. The plished that the inverse Fourier transforms with respect to αν
vicinity of the branch point should be treated carefully. To en- no longer have to be evaluated for each x µ separately, which
sure that the branch point is out of the propagation regime is the key simplification of the GS propagator compared with
within the thin slab, we have chosen c0 smaller than the min- the “exact” thin-slab propagator (g). We obtain
imum medium wave speed within the slab, as illustrated in
Figure 1. Unlike the GS expansion, the split-step Fourier
method does not suffer from the presence of a branch point. g (±) (xµ , x3 ; xυ0 , x30 ) ' g 0(±) (xµ , x3 ; xυ0 , x30 )
The choice of a background wave speed c0 greater than the
minimum wave speed in each thin slab is therefore possi- + g 1(±) (xµ , x3 ; xυ0 , x30 ), (25)
ble for this approximation. One may then choose, for exam-
ple, the mean or the median of the medium velocities in the
with
slab as a reference. Different choices yield different opera-
tors. It is not obvious how to find the “optimum” background
£ ¤
medium. g 0(±) (xµ , x3 ; xυ0 , x30 ) = exp ∓sγ11 (xµ , x̄3 ; 0) 1x3
In the phase-screen approximation [cf. equation (19)], the Z
Taylor expansion of 1/γ 0 in αν αν was truncated after the first
× (s/2π )2 dα1 dα2 exp[−is ασ (xσ − xσ0 )]
term. The accuracy of this approximation is increased by simply
adding more terms, e.g., £ ¤
× exp ∓sγ 0 (x̄3 ; αν ) 1x3 , (26)
γ1 (xµ , x3 ; αν ) ' γ 0 (x3 ; αν ) + 12 u(xµ , x3 )
h ¡ ¢3 i
c0 (x3 ) − 12 αν αν c0 (x3 ) , (21) and

£ ¤
yielding an angular correction in the vertical slowness. The ac- g 1(±) (xµ , x3 ; xυ0 , x30 ) = exp ∓sγ11 (xµ , x̄3 ; 0) 1x3
curacy however, is bounded by the accuracy of the first-order Z
GS expansion (De Hoop et al., 2000, their Figure 3).
× (s/2π )2 dα1 dα2 exp[−is ασ (xσ − xσ0 )]
£ ¤
The scalar generalized-screen propagator × exp ∓sγ 0 (x̄3 ; αν ) 1x3
£ ¡ ¢¤
With the analysis given in equation (16), we write equation ×1x3 ∓s γ11 (xµ , x̄3 ; αν ) − γ11 (xµ , x̄3 ; 0) . (27)
(10) in the form
Z
g (±) (xµ , x3 ; xυ0 , x30 ) ' (s/2π)2 dα1 dα2 Since

× exp[−is ασ (xσ − xσ0 )] X



[u(xµ , x̄3 )] j
γ11 (xµ , x̄3 ; 0) ∼ aj £ ¤2 j−1
£ © ª ¤ γ 0 (x̄3 ; 0)
× exp ∓s γ 0 (x̄3 ; αν ) + γ11 (xµ , x̄3 ; αν ) 1x3 , (22) j=1
£ ¤−1 £ 0 ¤−1
where
= c(xµ , x̄3 ) − c (xµ , x̄3 ) ,

x̄3 = x3 − 12 1x3 = x30 + 12 1x3 . (23) which is a spatial phase correction term that may be found in
the split-step Fourier method (Stoffa et al., 1990), we obtain
Expanding the exponential, we seek an approximation of the
propagator that matches the structure of the GS expansion. We
therefore expand the perturbation term about vertical propa- g 0(±) (xµ , x3 ; xυ0 , x30 ) =
gation and write £ ¡£ ¤−1 £ 0 ¤−1 ¢ ¤
exp ∓s c(xµ , x̄3 ) − c (x̄3 ) 1x3
γ11 (xµ , x̄3 ; αν ) = γ11 (xµ , x̄3 ; 0) Z
£ ¤
+ γ11 (xµ , x̄3 ; αν ) − γ11 (xµ , x̄3 ; 0) , × (s/2π )2 dα1 dα2 exp[−is ασ (xσ − xσ0 )]
£ ¤
and × exp ∓sγ 0 (x̄3 ; αν ) 1x3 . (28)
£ ¡ ¢ ¤
exp ∓s γ11 (xµ , x̄3 ; αν ) − γ11 (xµ , x̄3 ; 0) 1x3
£ ¤ The leading-order constituent term of the GS propagator,
' 1 ∓ s γ11 (xµ , x̄3 ; αν ) − γ11 (xµ , x̄3 ; 0) 1x3 . (24) g 0(±) , is the propagator of the split-step Fourier method, i.e.,
a propagator associated with the background medium for the
Expanding the exponential separates dependencies of the thin slab and a phase correction term in space that is accurate
propagator on xµ and on αν , which allows spatial dependencies for the vertical propagation only. The higher order term in the
to be taken out of the Fourier integral (22). We have accom- GS propagator then follows as
1556 Le Rousseau and de Hoop

g 1(±) (xµ , x3 ; xυ0 , x30 ) = the division being carried out in some stable sense, and the
normalizing operator N being given by (De Hoop et al., 2000)
£ ¡ £ ¤−1 ¢ ¤ ¯ ¯ · ¸
exp ∓s [c(xµ , x̄3 )]−1 − c0 (x̄3 ) 1x3 ¯ p ¯¯−1 p
Z N [1 + p + iq] = exp(iq) ¯¯1 + 1+ .
1 + iq ¯ 1 + iq
× (s/2π)2 dα1 dα2 exp[−is ασ (xσ − xσ0 )]
In effect, this normalization corrects for the error introduced
£ ¤ by the expansion of the exponential in equation (24) and re-
× exp ∓sγ (x̄3 ; αν ) 1x3
0
stores the amplitude behavior exactly in the case of a constant
£ ¡ ¢¤ perturbation u (i.e., a medium with constant wave speed higher
×1x3 ∓s γ11 (xµ , x̄3 ; αν ) − γ11 (xµ , x̄3 ; 0) , (29)
than the reference wave speed).
which, upon replacing γ11 by its truncated expression as in equa-
THE SCALAR GENERALIZED-SCREEN ALGORITHM
tion (16), gives
Here, we discuss the GS algorithm based upon equations
g 1(±) (xµ , x3 ; xυ0 , x30 ) =
(22)–(30). We denote the (one-way) wavefield by W , and carry
£ ¡£ ¤−1 £ 0 ¤−1 ¢ ¤ out the wave propagation in the frequency domain, with each
exp ∓s c(xµ , x̄3 ) − c (x̄3 ) 1x3 frequency component computed independently. The down-
( Z ward continuation for modeling and imaging with the one-way
X
n
× (∓)s1x3 a j [u(xµ , x̄3 )] j
(s/2π )2 dα1 dα2 propagator is performed according to the decomposition of the
j=1 vertical slowness symbol into one background term and a series
of perturbation terms as in equation (16).
£ ¤ Let the current depth be set to x30 = z, and set x̄3 = x30 + 12 1x3
× exp[−is ασ (xσ − xσ0 )] exp ∓sγ 0 (x̄3 ; αν ) 1x3
as before. Following equation (30), we introduce the interme-
" #) diate field quantities w0 , . . . , wn according to (step 1)
1 1
× £ ¤2 j−1 − £ ¤2 j−1 . (30)
γ 0 (x̄3 ; αν ) γ 0 (x̄3 ; 0) w0 (xµ , s) =
£ ¡ £ ¤−1 ¢¤
exp −s 1x3 [c(xµ , x̄3 )]−1 − c0 (x̄3 ) W (xµ , x30 , s),
The second constituent term of the GS propagator, g 1(±) , arises
from the higher order terms of the GS expansion. We call n the w1 (xµ , s) = −s 1x3 a1 u(xµ , x̄3 ) w0 (xµ , s),
order of the GS approximation. As mentioned, the higher the
order, the higher the accuracy for wide-angle propagation. w2 (xµ , s) = −s 1x3 a2 u 2 (xµ , x̄3 ) w0 (xµ , s),
The split-step Fourier propagator simply yields a shut- ..
tling between the frequency-horizontal space and frequency-
.
horizontal slowness domains and a multiplication in each do- wn (xµ , s) = −s 1x3 an u n (xµ , x̄3 ) w0 (xµ , s).
main. Each additional term of the GS expansion (16) requires
an additional inverse Fourier transform in space as expressed in The quantity w0 represents the split-step Fourier term; the
equation (30). As the computational complexity of the down- higher-order terms increase the accuracy for wider-angle prop-
ward continuation in the split-step Fourier method is propor- agation. The intermediate field quantities are then Fourier
tional to 2N1 N2 log2 N1 N2 (Nµ denoting the numbers of sam- transformed to the horizontal-wavenumber domain (step 2):
ples in the xµ -direction), the complexity of our nth-order GS Z
approach is proportional to (2 + n)N1 N2 log2 N1 N2 . w̃ j (αν , s) = w j (xµ , s) exp[−ixσ kσ ] dx1 , dx2 , (31)
Z
= w j (xµ , s) exp[isxσ ασ ] dx1 , dx2 , j = 0, . . . , n.
Normalization of the scalar generalized-screen propagator
The Taylor expansion of the exponential in equation (24) de- The wavefield at depth x3 = z + 1x3 then follows as
stroys the unitarity of propagator (8) and, hence, the amplitude £ ¤
characteristics of the propagator. Depending on the wavenum- W̃ (x3 ; αν , s) = w̃0 (αν , s) exp −s1x3 γ 0 (x̄3 ; αν )
ber, or the local angle of propagation, part of the energy may " Ã !
be amplified or attenuated incorrectly. It can also lead to a nu- w̃1 (αν , s) 1
×N 1 + − c (x̄3 ) + · · ·
0
merical algorithm that is unstable over a large range of down- w̃0 (αν , s) γ 0 (x̄3 ; αν )
ward continuation. To restore, approximately, the amplitude à !#
behavior, we apply a normalizing operator and obtain the GS w̃n (αν , s) 1 £ 0 ¤2n−1
propagator gG S P : + £ ¤2n−1 − c (x̄3 )
w̃0 (αν , s) γ 0 (x̄3 ; αν
gG S P (xµ , x3 ; xυ0 , x30 ) = g 0(±) (xµ , x3 ; xυ0 , x30 )
(±)
(step 3). Finally, we carry out the inverse Fourier transform
" # W̃ (x3 ; αν , s) → W (xµ , x3 ; s) (step 4).
g 1(±) (xµ , x3 ; xυ0 , x30 ) The sequence of steps illustrates the shuttling between the
× N 1 + 0(±) , frequency-horizontal space domain and frequency-horizontal
g (xµ , x3 ; xυ0 , x30 )
wavenumber domain that is performed with the GS approach.
Generalized Screens 1557

The method departs from the exact solution through the horizontal propagation is due to the choice of c0 : c0 is only
(approximate) expansion of the vertical slowness operator. two-thirds of the actual wave speed c.
The case of a constant medium perturbation provides in- Figure 3 compares 3-D computations for the first four orders
sight in how wavefronts evolve based on Huygens’ principle of the GS propagators and for the split-step Fourier method.
(Huygens, 1690; Hörmander, 1983). Let the background Again, the dashed curves show the true wavefront location.
medium be characterized by a wave speed c0 that is two-thirds From Figure 3, we conjecture that, as a rule of thumb, the
of the true wave speed. The top of Figure 2 shows (instanta- split-step Fourier method is accurate up to 17◦ , the first-order
neous) wavefronts constructed as the polar reciprocal of the GS is accurate up to 34◦ , the second-order GS up to 48◦ , the
local slowness surface shown in Figure 1. In the bottom part third-order GS up to 55◦ , and the fourth-order GS up to 62◦ ,
of Figure 2, we show numerical impulse responses of the one- when such a perturbation occurs (the background wave speed
way propagator for different orders of generalized screens as being only two-thirds of the actual wave speed). The rate of
well as for the split-step Fourier method. For a constant per- convergence with order is also suggestive in Figure 3. As in any
turbation, the predicted wavefronts and the actual computed Taylor expansion, the first terms in the GS expansion bring the
wavefronts can be overlain exactly. The exact response is plot- main contribution to the operator, and the higher-order terms
ted dashed in the top part of the figure. Note that the accuracy have lesser contributions, slowly increasing the accuracy.
varies with propagation angle (or migration dip upon imaging),
and that this accuracy varies with (local) medium perturbation.
In all the approximations, independent of order n, the propa- BRANCH POINTS
gation speed in the horizontal directions approaches c0 , which The terms in the GS expansion contain powers of 1/γ 0 in-
causes any approximate wavefront to fold inwards away from creasing with order [cf. equation (16)]. These powers induce
the true wavefront. Indeed, the normal to any slowness surface branch points at αν αν = −[c0 ]−2 , corresponding to grazing
in Figure 1 is horizontal or converges to horizontal when the propagation in the background medium (cf. Figure 1). The
horizontal slowness is 1/c0 . As such, the GS approximation dif- branch points also appear in expansion (30) for the GS propa-
fers, for example, from the paraxial approximation where the gator.
accuracy with dip is independent of the medium (De Hoop and For a constant-perturbation medium, the normalization
De Hoop, 1992). The GS inaccuracy at horizontal propagation operator exactly compensates the errors arising from the
is inherited from the GS expansion of the vertical slowness
symbol. Note, however, that for nonhorizontal propagation,
including very large angle propagation, any accuracy can be
obtained depending on the order of the GS approximation
chosen. We should also note that the large error seen here for

FIG. 2. Wavefronts in a constant-perturbation medium associ-


ated with the various generalized-screen approximations: sec- FIG. 3. Three-dimensional wavefield snapshots in a constant-
ond-order (GSP2), first-order (GSP1), and split-step Fourier. perturbation medium associated with the various generalized
Top: as calculated as polar reciprocal of the slowness surface screen approximations: fourth-order (GSP4), third-order
(the exact and background wavefronts are shown dashed); bot- (GSP3), second-order (GSP2), first-order (GSP1), and split-
tom: numerical wavefront. step Fourier. The exact wavefront is shown dashed.
1558 Le Rousseau and de Hoop

expansion of the exponential in equation (24). Yet, in later- the path should depart from the real axis in the vicinity of the
ally heterogeneous media such as the Marmousi model (cf. branch points; thus, we use different deformations for the dif-
Figure 4), these approximations can create errors that can ferent terms in the GS expansions. On the other hand, such
be boosted by large values of 1/γ 0 close to a branch point. incomplete contour deformation does affect the source signa-
Figure 6a shows a snapshot of the wave field, computed with ture. This enforces us to keep the contours as close as possi-
the second-order GS algorithm, in the region of the Marmousi ble to the real axis. Hence, the precise choice of contours is a
model shown in Figure 4 when the wavenumber is allowed to compromise.
get close to the branch point. The source is located at the star in
Figure 4. The result is noisy; the noise is associated with high- ACCURACY ANALYSIS
angle propagation (i.e., when wavenumbers reach the vicinity
We illustrate the accuracy of the GS algorithm for two sce-
of the branch point).
narios: the 2-D Marmousi model and the SEG-EAGE 3-D salt
In equation (30), the path of integration in the α1 -plane
model. Our analysis emphasizes numerical modeling. We gen-
should be chosen appropriately around the branch cuts. To
erate Green’s functions and focus our observations on second-
avoid getting too close to the artificial branch points, we use
arrival energy and multipathing. Imaging invokes an averag-
contour deformation in the complex plane, as illustrated in
ing (stacking) process. To learn about the prestack migration
Figure 5. Comparing the result in Figure 6a with the case in
operator, we hence prefer to analyze modeling instead. We
Figure 6b, where the contour deformation has been applied,
shall, however, show depth-migration results in the Marmousi
note the removal of the noise associated with high-angle prop-
model.
agating energy.
In practice, all factors appearing in the GS algorithm are
Modeling
computed on the actual deformed contour except for the inter-
mediate field quantities in the wavenumber domain, which are Marmousi model.—The medium complexity and the maxi-
evaluated on the real axis. The latter quantities are assumed to mum propagation angle determine the sophistication of the ap-
vary smoothly in the vicinity of the real axis and hence are ap- proximation required. To illustrate this, we consider a region of
proximated by their values on the real axis as evaluated with the the Institut Français du Pétrole (IFP)’s Marmousi model that
(fast) Fourier transforms. We have made a zero-order Taylor
approximation where the contour departs from the real axis.
On the one hand, the higher the power of [γ 0 ]−1 , the further

FIG. 4. Part of the Marmousi wave speed model used to gener-


ate the snapshots of Figures 6–12. The star locates the position
of the source.

FIG. 6. Snapshot of the wavefield at time t = 0.95 s computed


with the second order of the generalized-screen method: (a) no
contour deformation applied, (b) contour deformation applied
as in Figure 5. Figure 4 shows the wave speed model used and
FIG. 5. Applied contour deformation. the source location.
Generalized Screens 1559

includes the target reservoir (Figure 4). A point source (star) reference wave speed. Use of the median [ 12 (cmax − cmin )] for
is located at the reservoir horizon, below a complex part of the reference wave speed would give a poor result (not shown
the model (anticline, unconformity, faults) that exhibits signif- here). Use of the same reference wave speed as for the GS
icant horizontal wave speed variations. Being based on actual method (i.e., the minimum wave speed in the thin slab) yields
geology (Bourgeois et al., 1991), the Marmousi model intro- even poorer result because it increases the contrast between
duces complexities that can be encountered in practice. The the true medium and the background medium.
source is excited at time t = 0, and the pressure field is imaged With the number of reference velocities chosen for the PSPI
at time t = 0.95 s, creating the snapshots in Figures 7–10. For method, we guaranteed a better accuracy, but at higher cost,
comparison, we modeled the two-way wavefield with the full than that of the split-step Fourier method. To compare algo-
acoustic wave equation (finite-difference time-domain, second rithm accuracies, we look at the degree at which the wave-
order in space, second order in time) and shall use that wave- front is curved downward or upward with respect to the full-
field (Figure 7) as a reference ( the computational complexity acoustic wave-equation generated wavefront. The split-step
of the full acoustic wave equation method in two dimensions, Fourier (Figure 8) result gives an impulse response with a
and even more so in three dimensions, compares unfavorably wider shape (upward curved) for the first arrival than the
with the fast algorithms developed in this paper). In modeling full-wave result, whereas the PSPI method (Figure 9) gives
the upgoing wavefield, we compare the GS method with the a correct general shape for the first arrival. Neither of these
split-step Fourier and the phase-shift-plus-interpolation (PSPI) methods, however, gives accurate images of the second-arrival
methods. In the PSPI method (Gazdag and Sguazzero, 1984), energy (although the PSPI method still yields a better result).
wave components are upward or downward continued as if the This second-arrival energy results from multipathing associ-
medium were laterally homogeneous but using several refer- ated with the complexity and the heterogeneity of the Mar-
ence velocities to accommodate lateral wave speed variations: mousi model. The multipathing here occurs with scattering
in the frequency-space domain, at any point in depth, an inter- at reasonably wide angles. For the split-step Fourier method,
polation is performed using two fields associated with the two the limited accuracy for wide-angle propagation explains the
adjacent reference velocities. poor modeling for the second-arrival energy. For the PSPI
Unlike the GS method, the split-step Fourier method does method, inaccuracies are attributable to the linear interpola-
not suffer from the presence of a branch point. We can there- tion scheme.
fore use a background wave speed c0 that is greater than the To begin with the results of the various orders of the GS
minimum wave speed in each thin slab. In Figure 8, we use the method, the second-order algorithm (Figure 6b) creates a wave
arithmetic mean of the wave speed across the thin slab as a front whose shape, as expected, is slightly curved downward
with respect to the full-acoustic wave-equation wavefront. As
the order of the GS expansion increases, the GS wavefront
matches the true one more closely. The wavefront of the fourth-
order solution (Figure 10) almost coincides with the true wave-
front. With the second-order GS method (Figure 6b), one can
already see good definition of the second-arrival energy due to
a higher accuracy for wide-angle propagation than that in the
split-step Fourier method. The better positioning and modeling
of the later-arrival energy constitutes a key contribution to the
imaging of complex structures such as those in the Marmousi
model where these arrivals carry a significant part of the en-
ergy. Observe that subtle differences in Figures 8–10 represent
differences in positioning of the order of a 100 m.
To highlight the conservation of source signature for the
FIG. 7. Snapshot of the wavefield at time t = 0.95 s with the different orders of the GS expansion—under the incomplete
full two-way acoustic wave equation. contour deformation—and the finite-difference, PSPI, and

FIG. 8. Snapshot of the wavefield at time t = 0.95 s with the FIG. 9. Snapshot of the wavefield at time t = 0.95 s with the
split-step Fourier method. PSPI method.
1560 Le Rousseau and de Hoop

splitstep Fourier approaches, we extract a single trace on the the Bremmer series (De Hoop, 1996; Van Stralen et al., 1998;
far left out of the snapshots shown in Figure 7, 8, 6b, 10, and 9. De Hoop et al., 2000), which necessitates the introduction of a
These traces are combined in Figure 11. Observe that the “ar- GS representation of the reflection kernel (Appendix C). It is
rival” of the fourth-order GS and PSPI methods align with the illustrated in Figure 12, where the fourth-order GS algorithm
one obtained with the finite-difference computation, unlike the is used to calculate the first three terms of the Bremmer series.
ones obtained with the split-step Fourier and the second-order The first term corresponds to the upcoming energy as given by
GS methods. At these propagation angles, the fourth-order GS the one-way operator. The second term corresponds to energy
approximation suffices, whereas the lower orders would not. that first travels upward, then downward. The third term corre-
The source signature, in particular in the fourth-order GS ap- sponds to energy that travels upward, downward, and upward
proximation, is clearly preserved at the expense of creating a again. Here, we make use of the first-order GS representation
weak precursor associated with the compromise in contour de- of the reflection operator as described in Appendix C. Note
formation. Such precursor is also visible in Figures 6b and 10. that some of the multiple events found in Figure 7 are miss-
Furthermore, when taking multiples into account, wide- ing because they would belong to yet higher-order terms in
angle propagation can be particularly significant, as mentioned the Bremmer series. Note in Figure 12 that some branches of
before. Incorporating multiples is achieved with the use of the wavefronts do propagate almost horizontally and accurate
results require higher order GS.

SEG-EAGE 3-D salt model.—For a 3-D modeling demon-


stration, we use the SEG-EAGE 3-D salt model. Significant
multipathing occurs in this model, and we shall illustrate how
well the GS propagator accounts for this. Figure 13 shows two
vertical profiles of wave speed across the SEG-EAGE 3-D salt
model, one parallel to the so-called in-line direction, and the
second one parallel to the orthogonal cross-line direction. The
profiles intersect at the center of the model. As in the Marmousi
model, we place a point source in the zone of interest (i.e., be-
neath the salt body). The source location is represented by an

FIG. 10. snapshot of the wavefield at time t = 0.95 s with the


fourth order of the generalized-screen method.

FIG. 12. Snapshot of the wavefield at time t = 0.95 s with the


fourth order of the generalized-screen modeling the three first
terms of the Bremmer series.

FIG. 11. First trace on the left of the snapshots shown in Figures
7, 8, 6b, 10, and 9: (1) finite difference, (2) split-step Fourier, FIG. 13. Vertical profiles across the center of the SEG-EAGE
(3) second-order generalized-screen, (4) fourth-order general- salt model. The asterisk locates the position of the source; the
ized-screen, (5) PSPI. dashed line indicates the region detailed in Figures 14–18.
Generalized Screens 1561

asterisk in Figure 13. Note the stairstep shape of the base of the Imaging
salt body. We expect this artificial roughness, which is due to the
coarse sampling of the model in depth, to create diffractions Laterally varying medium.—The difference in accuracy
that will appear as noise in the wavefield. between the split-step Fourier method and the GS approach
In our example, the source is excited at time t = 0, and we is illustrated with a 2-D imaging experiment (Figure 19). The
use the second-order GS algorithm to image the field at time 2D section is composed of various reflectors with dips ranging
t = 1.2 s (snapshots in Figures 14–18). Both the in-line section from 0◦ to 75◦ . The wave speed profile is characterized by
in Figure 14 and the cross-line section in Figure 15 show that a gradient with a horizontal component of 0.1 s−1 and a
second arrivals carry a significant part of the upgoing energy.
They both show that, with the geometry of the salt body, partial
waveguiding can occur, which could imply that a significant part
of information is contained at large offsets in the scattered field
recorded at the surface. Note the occurrence of a triplication
in the cross-line section in Figure 15. The various horizontal
sections of the 3-D image (Figures 16–18) illustrate the strong
imprint that the salt structure imposes on the wavefront. The
deepest slices (Figure 16) intersect the salt body, which explains
the higher noise level also visible in the vertical sections; this
noise, again, corresponds to diffracted energy from the rough
salt bottom.

FIG. 16. Snapshot of the wavefield at time t = 1.2 s with the sec-
ond order of the generalized-screen method: horizontal section
at depth 1500 m. the dashed lines represent the vertical sections
as shown in Figures 14 and 15.

FIG. 14. Snapshot of the wavefield at time t = 1.2 s with the


second order of the generalized-screen method: in-line section.
The three horizontal dashed lines represent the depth location
of the horizontal sections of Figures 16–18.

FIG. 17. Snapshot of the wavefield at time t = 1.2 s with the


FIG. 15. Snapshot of the wavefield at time t = 1.2 s with second order of the generalized-screen method: horizontal sec-
the second order of the generalized-screen method: cross-line tion at depth 1000 m. The dashed lines represent the vertical
section. sections as shown in Figures 14 and 15.
1562 Le Rousseau and de Hoop

vertical component of 0.4 s−1 . Figure 19 shows results after associated with the steepest events, is better handled. Note
migrations of a zero-offset section (not shown) with the that the steepest event appears weaker on the GS generated
split-step Fourier algorithm and the fourth-order GS method. section than on the split-step Fourier section. This is due to
In the split-step Fourier result, reflectors steeper than about the proximity of the branch point to the propagation angles
45◦ are mispositioned with an error that grows with dip. associated with the steepest reflector dip. The cascade of the
Again, the accuracy with dip is a function of the lateral expansion of the thin-slab propagator and the normalization
medium variation for both the split-step Fourier method and are progressively less accurate for the phase as one departs
the GS approach. In the GS generated section, reflectors are from vertical propagation and constant medium perturbation.
accurately positioned because the wide-angle propagation, This introduces some amplitude inaccuracy.

Application to prestack depth migration.—To illustrate the


accuracy of the different GS expansions for imaging, we in-
corporated the GS algorithm in prestack depth migration of
shot gathers using the entire Marmousi data set (Figure 20). In
the migration procedure, the fields associated with the source
and receivers are downward continued. For each point in
the subsurface, the two fields are correlated at zero time lag
(Claerbout, 1986).
Figures 21a and 21b compare the images obtained with the
split-step Fourier version of the phase-screen method and the
second-order GS approach. Because of the two additional cor-
rection terms appearing in the expansion, dipping events are
more accurately positioned and better focused by the GS ap-
proach. Regions where significant differences occur are high-
lighted in Figures 21a and 21b. The anticline (above the reser-
voir) of the Marmousi model poses a challenge to imaging for
any migration algorithm because of multipathing and wide-
angle scattering. The GS method accommodates these phe-
nomena better than does the split-step method and yields more
continuity in, and less deformation of, the reflectors.

DISCUSSION

FIG. 18. Snapshot of the wavefield at time t = 1.2 s with the sec- The GS wave extrapolation method is based on the decom-
ond order of the generalized-screen method: horizontal section position of the medium into a background component and
at depth 500 m. The dashed lines represent the vertical sections a perturbation. In lowest order, the GS approximation may
as shown in Figures 14 and 15. be simplified to yield the phase-screen and split-step Fourier
methods. The GS method extends these two methods and
can accommodate more significant and rapid horizontal wave
speed variations. In the process of migration, the GS method
provides a choice of algorithm suitable for the complexity of a
particular medium.
With the GS representation of the propagator is associated
a vertical slowness symbol that generalizes the phase-screen
and split-step Fourier symbol and has the dependencies on the
spatial coordinates and the wavenumber factorized, inducing
the structure of the GS propagator. The GS expansion of the
vertical slowness symbol is justified in De Hoop et al. (2000).
The GS expansion provides a fundamental simplification of
the one-way propagator; it does not have to be evaluated at

FIG. 19. Top: model having reflectors with dips of 0◦ , 215◦ ,


30◦ , 45◦ , 60◦ , and 75◦ . Middle: migration of the modeled data
with split-step Fourier method. Bottom: migration of the
fourth-order generalized-screen method (GSP 4). FIG. 20. Marmousi wave speed model.
Generalized Screens 1563

every output point. We obtained an algorithm that works as We have compared the GS method with some of its com-
a shuttling between the frequency-horizontal space domain peting algorithms. All these algorithms account up to a certain
and the frequency-horizontal wavenumber domain. Each degree for multipathing: one-way (split-step Fourier, PSPI),
additional term in the GS expansion increases the accuracy. two-way (finite differences). Another approach to approx-
This amounts to an additional Fourier transform for each imate the one-way propagator (10) is the filter approach.
additional order in the expansion. In anisotropic media such Like our approach, the filter approach aims at approximat-
as transversely isotropic media with a vertical symmetry axis, ing the propagation of singularities. Unlike our approach, it
the influence of the anisotropic parameters in the vertical does not yield a comprehensive recursion that generates in-
slowness symbol occurs only at nonvertical propagation. creasingly accurate one-way propagators on and away from
Thus, the split-step-Fourier type approximation (Stoffa et al., the wavefront set. The filter approach was introduced by
1990) does not have the degree of freedom to account for Holberg (1988) and Blacquière et al. (1989), and extended
anisotropy or to shape the slowness surface appropriately. The to three dimensions by Hale (1991a, b). In three dimensions,
GS method can be extended to these media, as shown in the this approach is based on the construction of a 1-D finite-
companion paper (Le Rousseau and De Hoop, 2001). length P filter with coefficients h n and
P Nhspectrum J (cos k10 ) =
Nh −1 −1
We illustrated the GS propagator’s accuracy primarily h 0 + 2 n=1 h n cos(k1 n) = h 0 + 2 n=1 Tn (cos k10 ), where Tn
0

through modeling. The mathematical accuracy analysis of the denote Chebychev polynomials (k10 is the wave number in the
GS approximation was carried out by De Hoop et al. (2000). 1-direction scaled by the sampling rate, i.e., k10 = −sα1 1x1 ),
Here, we focused on its numerical aspects. We chose two extended to two dimensions q when combined with an approx-
synthetic models, both representative of the real world: the imation of cos k 0 = cos k102 + k202 ' G(k10 , k20 ) = −1 + 12 (1 +
Marmousi model (offshore Africa) and the SEG/EAGE 3-D cos k10 )(1 + cos k20 ) proposed by McClellan (1973) to obtain an
salt model (Gulf of Mexico). Each of these models provided approximate circular filter and later improved by Hale (1991a).
insight into the capacity of the GS method to perform wave We associate the vertical symbol
extrapolation, and showed how it compares with the split-step
Fourier method, as well as PSPI and full-wave time-domain 1
γ H (xµ , αν ) = − log J (G(−s α1 1x1 , −s α2 1x2 ))
finite-difference methods. The GS modeling capacity is illus- s1x3
trated in both two and three dimensions. In this paper, we focus
on the computation of Green’s functions rather than on imag- with 1x2 = 1x1 , (32)
ing. Imaging invokes an averaging procedure that hides propa-
gation inaccuracies of the various methods. For completeness, with the filter, where the dependency on the transverse di-
however, we have shown some prestack imaging results in two rections, xµ , is contained in the coefficients h n . We observe
dimensions. that this dispersion relation which characterizes the propa-
gator is dependent on the frequency and the spatial sam-
pling rate. Note also that the use of a circular filter requires
the same sampling rate in the in-line and cross-line di-
rections (different sampling rates would require the design
of an elliptic filter). In Figures 22a and 22b, we illustrate both
the accuracy of the filter and the accuracy of the McClellan
transformation. In Figure 22, grazing incidence corresponds to
0.8 Nyquist and one can see the degradation of the accuracy
of the McClellan transformation at large propagation angle.
In this example, we use all the coefficients h n to match the
thin-slab propagator. In practice, this cannot be done because
it implies a nonunitary propagator. Hale (1991b) solved this
problem, decreasing the amplitude smoothly to zero beyond
half Nyquist. This impacts the propagation of energy at wide
angle, in view of the cascade of extrapolations. Note that in
Figures 22a and 22b, the vertical slownesses become complex
where their real parts depart from the exact one. Figure 22c
shows the computation of the wavefront sets associated with
the vertical slownesses shown in Figures 22a and 22b. The wave-
front sets truncate where the vertical slownesses become com-
plex, which does not happen in our GS approach. Concerning
the filter-McClellan approach, the following can be said: (1)
the filter will not be exact even in an (almost) homogeneous
medium, (2) the accuracy of the filter will depend on the sam-
pling rate which for fixed frequency will depend on the local
medium properties (i.e., the accuracy will vary with location),
(3) the same remarks apply to the McClellan transformation,
FIG. 21. Prestack depth migration of the Marmousi data and (4) intrinsic instability enforces dense sampling to “open
(a) with the split-step Fourier method, and (b) with the sec-
ond-order generalized-screen method. Highlighted are fea- up” the wavefront set. These features are avoided in the GS
tures where clear differences can be observed between the two method. As far as the computational complexity is concerned,
sections. the filter-McClellan and GS methods compare as follows:
1564 Le Rousseau and de Hoop

geneities. Nowadays, 3-D surveys are common, and some areas


present geologic complexities where multipathing cannot be
ignored. Unlike traditional efficient asymptotic methods (such
as Kirchhoff) that use a single arrival (first or maximal energy),
the GS method can predict the effect of multipathing. The in-
troduction of the Bremmer coupling series yields a computa-
tion of the two-way Green’s functions. The GS-Bremmer ap-
proach yields a tool box that can be used for processing multi-
ples, velocity analysis, and inversion (De Hoop and De Hoop,
2000). The lowest order GS approximations has already found
application in seismic imaging (Huang and Wu, 1996; Wu and
Jin, 1997; Huang et al., 1999).
ACKNOWLEDGMENTS

The authors especially thank Elf Exploration Production for


financial support of this research. We thank Henri Calandra
for numerous discussions and Kees Wapenaar for his valuable
comments. This work was also supported by the members of
the Consortium Project on Seismic Inverse Methods for Com-
plex Structures at the Center for Wave Phenomena, Colorado
School of Mines.
REFERENCES
Berkhout, A. J., 1982, Seismic migration: Imaging of acoustic energy
by wave field extrapolation: Elsevier Science Publ. Co., Inc.
Blacquière, G., Debeye, H. W. J., Wapenaar, C. P. A., and Berkhout,
A. J., 1989, 3d table-driven migration: Geophys. Prosp., 37, 925–958.
Bourgeois, A., Bourget, M., Lailly, P., Poulet, M., Ricarte, P., and
Versteeg, R., 1991, Marmousi, model and data: Proc. 1990 EAEG
workshop on Practical Aspects of Seismic Data Inversion.
Bramley, E. N., 1977, The accuracy of computing ionospheric ra-
diowave scintillation by the thin-phase screen approximation: J.
Atmos. Terr. Phys., 39, 367–373.
Buckley, R., 1975, Diffraction by a random phase-changing screen: A
numerical experiment: J. Atmos. Terr. Phys., 37, 1431–1446.
Claerbout, J., 1986, Imaging the earth’s interior: Blackwell Scientific
Publications, Inc.
Cohen-Tannoudji, C., Diu, B., and Laloë, F., 1977, Mécanique quan-
tique, vol. 1: Hermann.
De Bruin, C. G. M., Wapenaar, C. P. A., and Berkhout, A. J., 1990,
Angle-dependent reflectivity by means of prestack migration: Geo-
physics, 55, 1223–1234.
De Hoop, M. V., and Brandsberg-Dahl, S., 2000, Maslov asymptotic
FIG. 22. (a) Real and (b) imaginary parts of the vertical extension of generalized Radon transform inversion in anisotropic
slownesses associated with the filter-McClellan approach. The elastic media: A least-squares approach: Inverse Problems, 16, 519–
562.
in-line curve correspond to the 1-D filter; the 45◦ curves cor- De Hoop, M. V., and De Hoop, A. T., 1992, Scalar space-time waves
respond to the 1-D filter combined (1) with McClellan’s and in their spectral-domain first- and second-order Thiele approxima-
(2) with Hale’s approximations of the circular filter at a 45◦ tions: Wave Motion, 15, 229–265.
azimuth (Hale, 1991a). The true vertical slowness is shown ——— 2000, Wavefield reciprocity and optimization in remote sensing:
dashed. In the present case grazing incidence corresponds to Proc. Roy. Soc. London, Ser. A, 456, 641–682.
0.8 Nyquist. (c) Wavefronts computed as polar reciprocals of De Hoop, M. V., Le Rousseau, J. H., and Wu, R.-S., 2000, Genera-
the slowness surfaces; the exact wavefront is shown dashed. lization of the phase-screen approximation for the scattering of
Note the truncation of the wavefronts. acoustic waves: Wave Motion, 31, 43–70.
De Hoop, M. V., 1996, Generalization of the Bremmer coupling series:
J. Math. Phys., 37, 3246–3282.
52Nh − 46 (Hale, 1991a)versus (2 + n) log2 N1 N2 . Typically, De Witte-Morette, C., Maheshwari, A., and Nelson, B., 1979, Path
integration in non-relativistic quantum mechanics: Physics Reports,
Nh = 20, n = 2, N1 = N2 = 103 ; thus, the filter-McClellan ap- 50, 255–372.
proach is less than one-tenth as efficient than the GS method. Feit, M. D., and Fleck, J. A., 1978, Light propagation in graded-
The GS method, with its (2 + n)N1 N2 log2 N1 N2 computa- index optical fibers: Appl. Opt., 17, 3990–3998.
Filice, J. P., 1984, Studies of the microscale density fluctuations in the
tional complexity (n is the order of the GS approximation and solar wind using the power law phase screen model: Ph.D. thesis,
Nµ are the numbers of samples in the xµ -direction), makes Univ. of Calif. San Diego.
feasible “wave equation” modeling and imaging in three di- Fishman, L., and McCoy, J. J., 1984a, Derivation and application of
extended parabolic wave theories I. The factorized Helmholtz equa-
mensions. Finite-difference methods extend to three dimen- tion: J. Math. Phys., 25, 285–296.
sions, but yield algorithms with greater computational com- ——— 1984b, Derivation and application of extended parabolic
wave theories II. Path integral representations: J. Math. Phys., 25,
plexity. In comparison, the computational complexity of the 297–308.
split-step Fourier method follows as 2N1 N2 log2 N1 N2 , whereas Fishman, L., de Hoop, M. V., and Van Stralen, M. J. N., 2000, Exact
the computational complexity of the PSPI approach is (1 + n ref ) construction of square-root Helmholtz operator symbols: The fo-
cusing quadratic profile: J. Math. Phys., 41, 4881–4938.
N1 N2 log2 N1 N2 + 2N1 N2 , where n ref is the number of refer- Fisk, M. D., and McCartor, G. D., 1991, The phase screen method for
ence velocities dependent on the degree of lateral hetero- vector elastic waves: J. Geophys. Res., 96, 5985–6010.
Generalized Screens 1565

Fisk, M. D., Charrette, E. E., and McCartor, G. D., 1992, A compa- McClellan, J. H., 1973, The design of two-dimensional digital filters
rison of phase screen and finite difference calculations for elastic by transformations: 7th Ann. Princeton Conf. on Inform. Sci. and
waves in random media: J. Geophys. Res., 97, 12409–12423. Syst., 247–251.
Flatté, S. M., Dashen, R., Munk, W. H., Watson, K. M., and Mercier, R. P., 1962, Diffraction by a screen causing large ran-
Zachariasen, F., 1979, Sound transmission through a fluctuating dom phase fluctuations: Proc. Cambridge Philos. Soc., 58, 382–
ocean: Cambridge Univ. Press. 400.
Gazdag, J., 1978, Wave equation migration with the phase-shift Ratcliffe, J. A., 1956, Some aspects of diffraction theory and their
method: Geophysics, 43, 1342–1351. application to the ionosphere: Rep. Prog. Phys., 19, 190–263.
Gazdag, J., and Sguazzero, P., 1984, Migration of seismic data by Schulman, L. S., 1981, Techniques and applications of path integration:
phase shift plus interpolation: Geophysics, 49, 124–131. John Wiley and Sons.
Hale, D., 1991a, Migration via McClellan transformations: Geo- Schwartz, L., 1966, Théorie des distributions, 2nd ed.: Hermann.
physics, 56, 1778–1785. Stoffa, P. L., Fokkema, J. T., de Luna Freire, R. M., and Kessinger,
——— 1991b, Stable explicit depth extrapolation of seismic wavefields: W. P., 1990, Split-step fourier migration: Geophysics, 55, 410–421.
Geophysics, 56, 1770–1777. Thomson, D. J., and Chapman, N. R., 1983, A wide-angle split-step
Holberg, O., 1988, Towards optimum one-way wave propagation: Geo- algorithm for the parabolic equation: J. Acoust. Soc. Am., 74, 1848–
phys. Prosp., 36, 99–114. 1854.
Hörmander, L., 1983, The analysis of linear partial differential opera- Treves, F., 1967, Topological vector spaces, distributions and kernels:
tor, vol. 1: Springer-Verlag. Academic Press.
——— 1985, The analysis of linear partial differential operator, vol. 3: ——— 1980, Introduction to pseudodifferential and Fourier integral
Springer-Verlag. operators, vol. 1: Plenum Press.
Huang, L.-J., and Wu, R.-S., 1996, 3d prestack depth migration with Van Stralen, M. J. N., De Hoop, M. V., and Blok, H., 1998,
acoustic pseudo-screen propagators: Mathematical Methods in Geo- Generalized Bremmer series with rational approximation for the
physical Imaging IV, Proc. SPIE 2822, 40–51. scattering of waves in inhomogeneous media: J. Acoust. Soc. Am.,
Huang, L.-J., Fehler, M. C., and Wu, R.-S., 1999, Extended local 104, 1943–1963.
Born Fourier migration method: Geophysics, 64, 1524–1534. Wapenaar, C. P. A., and Berkhout, A. J., 1989, Elastic wave field
Huygens, C., 1690, Traité de la lumière: Pierre van der Aa. extrapolation. Redatuming of single- and multi-component seismic
Knepp, D. L., 1983, Multiple phase screen calculation of the tem- data: Elsevier Science Publ. Co., Inc.
poral behavior of stochastic waves: Proc. IEEE, 71, 722–737. Wu, R.-S., and Jin, S., 1997, Windowed GSP (generalized screen propa-
Le Rousseau, J. H., and De Hoop, M. V., 2001, Scalar generalized- gators) migration applied to SEG-EAEG salt model data: 67th Ann.
screen algorithms in transversely isotropic media with a vertical sym- Internat. Mtg., Soc. Explor. Geophys., Expanded Abstracts, 1746–
metry axis: Geophysics, 66, 1538–1550, this issue. 1749.
Martin, J. M., and Flatté, S. M., 1988, Intensity images and statistics Wu, R.-S., 1994, Wide-angle elastic wave one-way propagation in het-
from numerical simulation of wave propagation in 3-D random me- erogeneous media and an elastic wave complex-screen method:
dia: Appl. Opt., 27, 2111–2125. J. Geophys. Res., 99, 751–766.

APPENDIX A
RELATIONSHIP BETWEEN ONE-WAY GREEN’S FUNCTION AND PROPAGATOR

The propagator, g (±) , associated with the one-way wave ator to expression (A-4) yields
equation satisfies the pseudodifferential equation
[∂3 ± s0] G (±) (xµ , x3 ; xυ0 , x30 )
[∂3 ± s0]g (±)
(xµ , x3 ; xυ0 , x30 ) = 0, (A-1)
= ±g (±) (xµ , x3 ; xυ0 , x30 ) ∂3 H (±[x3 − x30 ])
complemented with the initial condition
± H (±[x3 − x30 ]) [∂3 ± s0] g (±) (xµ , x3 ; xυ0 , x30 ). (A-5)
g (±)
(xµ , x3 = x30 ; xυ0 , x30 ) = δ(xµ − xµ0 ). (A-2)
The second term in equation (A-5) vanishes because of prop-
Let us consider a function φ(xµ , x3 ) that satisfies the one-way erty (A-1). Hence,
wave equation also. With properties (A-1) and (A-2), we have
Z [∂3 ± s0] G (±) (xµ , x3 ; xυ0 , x30 ) = ± g (±) (xµ , x3 ; xυ0 , x30 )
φ(xµ , x3 ) = dx10 dx20 g (±) (xµ , x3 ; xυ0 , x30 ) φ(xυ0 , x30 ), × [±δ(x3 − x30 )] = g (±) (xµ , x3 = x30 ; xυ0 , x30 ) δ(x3 − x30 )
(A-3) = δ(xµ − xµ0 ) δ(x3 − x30 ), (A-6)
hence the terminology “propagator”. We now prove that the which is the defining equation for the one-way Green’s func-
function G (±) (xµ , x3 ; xυ0 , x30 ), defined as tion. The occurrence of the Heaviside function forces the
Green’s function to be causal; the propagator is not.
G (±) (xµ , x3 ; xυ0 , x30 ) = ±H (±[x3 − x30 ]) In this paper, we define a one-way Green’s function with
× g (±) (xµ , x3 ; xυ0 , x30 ), (A-4) respect to the preferred spatial direction (x3 ). In the field of
quantum mechanics, the same arguments are used to define the
is a Green’s function associated with the one-way wave equa- one-way Green’s function in time using the propagator that sat-
tion. H is the Heaviside function. Applying the one-way oper- isfies the Schrödinger equation (Cohen-Tannoudji et al., 1977).

APPENDIX B
REDUCTION TO THE PRINCIPAL PARTS

Using the composition rule for symbols of pseudodifferential is transformed into an equation for the corresponding left sym-
operators (Treves, 1980), the operator equation, bols (De Hoop, 1996):
£ ¤
exp − i∂ασ0 Dxσ0 γ (xµ x3 ; ασ0 3 )γ (xσ0 x3 ; αν )|(xµ0 ,αν0 )=(xµ ,αν )
0 2 = A, (B-1) = a(xµ x3 ; αν ), (B-2)
1566 Le Rousseau and de Hoop

where γ is the left symbol of 0 and a is the left symbol of A, duces to the principal parts, γ1 and a2 , respectively, of the
i.e., symbols:
0(xµ , x3 ; Dν ) exp(isασ xσ ) ≡ γ (xµ , x3 ; αν ) exp(isασ xσ ),
γ1 γ1 = a2 = κρ + αν αν
A(xµ , x3 ; Dν ) exp(isασ xσ ) ≡ a(xµ , x3 ; αν ) exp(isασ xσ ).
= c−2 + αν αν . (B-4)
From the form of A in the acoustic-pressure-normalization
analog of equation (4) (De Hoop, 1996), we have
Referring to the contrast formulation made here and in
a = κρ + αν αν − ρ −1 (Dν ρ)iαν De Hoop et al. (2000), the “high-frequency” approximation
of equation (B-4) can also be expressed by assuming that the
− ρ −1 (Dν Dν ρ) − ρ −2 (Dν ρ)(Dν ρ). (B-3)
vertical-slowness symbol depends only on the magnitude (²)
As s → ∞, the composition of symbols tends to an ordi- of the medium perturbation, but not on the smoothness (Ä) of
nary multiplication, and the solution of equation (B-2) re- the medium perturbation, i.e. γ1 is O(Ä0 ).

APPENDIX C
THE GENERALIZED-SCREEN REPRESENTATION OF THE REFLECTION OPERATOR

The thin-slab propagator given in equation (8) is the one at x3 , the reflection symbol can be expressed as
associated with the one-way wave equation. To recover the ¡ ¢ ¡ ¢
γ1 xµ , x3+ ; αν − γ1 xµ , x3− ; αν
two-way wave propagation, one has to make use of the r1 (xµ , x3 ; αν )1x3 = ¡ ¢ ¡ ¢,
Bremmer series, as described by De Hoop (1996). The first γ1 xµ , x3+ ; αν + γ1 xµ , x3− ; αν
term of the series is the one-way wavefield, the second term is
(C-3)
the first (class one) multiple, the third term is the second mul-
tiple, etc. The evaluation of each term of the Bremmer series where
requires the use of the reflection operator given as
x3− = x3 − 12 1x3 ,
R = 12 0 −1 (∂3 0) (C-1) and

for constant density and a continuous medium. We associate x3+ = x3 + 12 1x3 .


to R its left symbol, r :
Upon introducing a background medium and a perturbation
Z Z on both sides of the interface, i.e.,
Rφ(xµ , x3 ) = (s/2π )2 dα1 dα2 dx10 dx20 r (xµ , x3 ; αν )
¡ ¢ £ ¡ ¢¤−2 £ 0 ¡ − ¢¤−2
u xµ , x3− = c xµ , x3− − c x3 ,
× exp[−is(xσ − xσ0 )ασ ] φ(xυ0 , x3 ).
and
In the framework of the present “high-frequency” approxima- ¡ ¢ £ ¡ ¢¤−2 £ 0 ¡ + ¢¤−2
tion, R reduces to its principal part R1 , with r reducing to its
u xµ , x3+ = c xµ , x3+ − c x3 ,
principal symbol r1 .
In a similar fashion as in Appendix B, one can show that such one can expand the reflection principal symbol with respect to
reduction to principal symbol, yields these two perturbations. In the GS expansion of the one-way
thin-slab propagator, the expansion of the exponential in equa-
∂3 γ1 tion (24) yields accuracy in the vicinity of vertical propagation.
r1 = . (C-2) We want to enforce such accuracy for the expansion of the
2γ1
reflection operator as well, and thus write
At a horizontal interface, the reflection symbol is simply the r1 (xµ , x3 ; αν ) = r1 (xµ , x3 ; 0)
reflection coefficient, but this is not true for a nonhorizontal
interface. For a nonhorizontal interface, the reflection of an + [r1 (xµ , x3 ; αν ) − r1 (xµ , x3 ; 0)]. (C-4)
incoming wave is split between the one-way propagation op-
We denote r10 (x3 ; αν ) the reflection symbol associated with
erator and the reflection operator. The reflection operator is
the upper and lower background media. We then expand
responsible for the interaction between the upgoing and down-
r1 (xµ , x3 ; αν ) − r1 (xµ , x3 ; 0) with respect to the perturbations
going constituents of the wavefield. The one-way wave opera-
u at x3− and x3+ , which, to first order, yields
tor describes the transverse scattering.
Like γ1 , the reflection symbol r1 can be expanded by in- r1 (xµ , x3 ; αν ) ' r1 (xµ , x3 ; 0) − r10 (x3 ; 0) + r10 (x3 ; αν )
troducing the same background medium and perturbation as
¡ ¢ ¡ ¢
within the thin slab through which propagation takes place. + a1− (x3 ; αν )u xµ , x3− + a1+ (x3 ; αν )u xµ , x3+ , (C-5)
Here, we present the GS expansion of the reflection symbol in
a medium sampled at discrete points. At an interface located with
Generalized Screens 1567

¡ ¢ £ 0 ¡ − ¢¤3
γ 0 x3+ , αν ³£ 0 ¡ + ¢¤2 £ 0 ¡ − ¢¤2 ´ ¡ −¢ 0¡ +¢ c x
− 0¡ − ¢ c x3 + c x3 + c x3 c x3 + 0 ¡ 3+ ¢
0
γ x3 , αν c x3
a1− (x3 ; αν ) = ¡ + − ¢
σ x , x , αν
à 3 ¡3 ¢!
£ 0 ¡ − ¢¤3 0 ¡ + ¢ γ 0 x3+ , αν
2αν αν c x3 c x3 1 + 0¡ − ¢
γ x3 , αν
+ ¡ ¢ , (C-6)
σ x3+ , x3− , αν

and
¡ ¢ £ 0 ¡ + ¢¤3
γ 0 x3− , αν ³£ 0 ¡ + ¢¤2 £ 0 ¡ − ¢¤2 ´ ¡ −¢ 0¡ +¢ c x
¡ + ¢ c x3 + c x3 − c x3 c x3 − 0 ¡ 3− ¢
0
γ x3 , αν
0 c x3
a1+ (x3 ; αν ) = ¡ + − ¢
σ x , x , αν
à 3 ¡3 ¢!
£ 0 ¡ + ¢¤3 0 ¡ − ¢ γ 0 x3− , αν
2αν αν c x3 c x3 1 + 0¡ + ¢
γ x3 , αν
− ¡ ¢ ,
σ x3+ , x3− , αν (C-7)

where
¡ ¢ £ ¡ ¢ ¡ ¢¤2
σ x3+ , x3− , αν = 1x3 γ 0 x3+ , αν + γ 0 x3− , αν
£ ¡ ¢ ¡ ¢¤2
× c0 x3+ + c0 x3− .

The superscript “−” refers to the upper medium and the su-
perscript “+” to the lower one, whereas the subscript refers to
the order in the GS expansion. We have
a1− (x3 ; 0) = 0, a1+ (x3 ; 0) = 0.
Figure C-1 illustrates the GS expansion of the real part of
the reflection symbol at an interface where the upper medium
is characterized by a constant perturbation and the lower
medium by a zero perturbation. In the top of the figure, the
exact reflection symbol (dashed) and the first two GS approx-
imations
√ are shown as a function of the horizontal slowness
p = −αν αν . The bottom part of Figure C-1 shows the relative
error of the first two GS approximations. Note the increasing
angular accuracy as the order of the GS expansion increases.
This accuracy will be required to predict multiple scattering at
large offset. However, in the vicinity of the critical angle, all
orders break down.
Note that to compute the nth term in the Bremmer series,
the reflection operator is only applied n times to the wavefield
while a cascade of thin-slab GS propagators is applied for each
of the terms. Such a cascade implies an accumulation of errors
associated with the inaccuracies of the thin-slab GS propagator FIG. C-1. Top: real part of the principal parts of the general-
(e.g., propagation very close to the horizontal direction). Such ized-screen reflection symbols: first- order (GSR1) and sec-
ond-order (GSR2) as a function of horizontal slowness p. The
accumulation of errors does not occur when applying the GS principal part of the exact symbol is shown dashed. Bottom:
reflection operator, since we usually consider the first few terms relative error of the first-order and second-order GS approxi-
in the Bremmer series only. mations.
1568 Le Rousseau and de Hoop

The GS representation (here, first order) of the principal yielding the reflection operator (Schwartz) kernel R1
part of the reflection operator then follows as (Schwartz, 1966; Treves, 1967; De Hoop et al., 2000):
Z Z ¡ ¢
(R1 φ) (xµ , x3 ) ' (s/2π )2 dα1 dα2 dx10 dx20 R1 (xµ , xυ0 , x3 ) ' r1 (xµ , x3 ; 0) − r10 (x3 ; 0) δ(xµ − xµ0 )
Z
× exp[−is(xσ − xσ0 )ασ ]
+ (s/2π)2 dα1 dα2 r10 (x3 ; αν )
£
× r1 (xµ , x3 ; 0) − r10 (x3 ; 0) + r10 (x3 ; αν )
¡ ¢ × exp[−is(xσ − xσ0 ) ασ ]
+ a1− (x3 ; αν )u xµ , x3−
¡ ¢¤ Z
¡ ¢
+ a1+ (x3 ; αν )u xµ , x3+ φ(xυ0 , x3 ), (C-8) + u xµ , x3− (s/2π )2 dα1 dα2 a1− (x3 ; αν )
which can be written as
¡ ¢ × exp[−is(xσ − xσ0 )ασ ]
(R1 φ) (xµ , x3 ) ' r1 (xµ , x3 ; 0) − r10 (x3 ; 0) φ(xµ , x3 ) Z
Z Z ¡ ¢
+ u xµ , x3+ (s/2π )2 dα1 dα2 a1+ (x3 ; αν )
+ (s/2π)2 dα1 dα2 dx10 dx20 r10 (x3 ; αν )

× exp[−is(xσ − xσ0 )ασ ] φ(xυ0 , x3 ) × exp[−is(xσ − xσ0 )ασ ]. (C-10)


Z Z
¡ ¢
+ u xµ , x3+ (s/2π)2 dα1 dα2 dx10 dx20 a1− (x3 ; αν ) The concept of reflection kernel was introduced by Berkhout
(1982, sections 4.6 and 6.2), Wapenaar and Berkhout (1989,
× exp[−is(xσ − xσ0 )ασ ] φ(xυ0 , x3 )
Z Z section III.3.3), and De Bruin et al. (1990). Equations (C-9)
¡ ¢
+ u xµ , x3+ (s/2π)2 dα1 dα2 dx10 dx20 a1+ (x3 ; αν ) and (C-10) can directly be cast into a numerical algorithm.
Each additional term in the GS expansion of the reflection
× exp[−is(xσ − xσ0 )ασ ] φ(xυ0 , x3 ), (C-9) operator requires two additional Fourier transforms.