Sie sind auf Seite 1von 33

Geophys. J . R. astr. SOC.(1976) 46, 201-233.

Exact and Approximate Generalized Ray Theory in


Vertically Inhomogeneous Media
C. H. Chapman
Department of Physics, University of Toronto, Toronto, Ontario, Canada
(Received 1976 March 5; in original form 1975 December 26)

Summary

The generalized ray method in a vertically inhomogeneous model is


formulated without any approximation by homogeneous layers. The
solution is obtained as an infinite series in multiply reflected waves.
Each term can be solved using the exact method or the plane-wave, firstmotion or geometrical approximations. It is shown that the firstmotion approximation of the series converges rapidly, the ratio of
successive terms in the infinite series being - (21+ 1)(21)(6/n).
In addition it is shown that the first-motion approximation, which
reduces to the geometrical approximation when the latter is valid, is a useful
alternative to geometrical ray theory, being more generally valid and
being almost as simple to compute.
1. Introduction

In a recent paper (Chapman 1974a), hereafter referred to as Paper I, the generalized


ray method was extended from models of homogeneous plane layers to inhomogeneous models where the density and wave velocities are functions of the vertical
co-ordinate. In this paper we develop this method more thoroughly.
Initially the generalized ray method was developed for problems with a single or no
interface (Cagniard 1939, 1962; Pekeris 1955a, b; de Hoop 1960). The exact impulse
response for each ray is obtained. The method can be extended to any ray in a model of
many homogeneous layers (Spencer 1960; Helmberger 1968; Muller 1970 and others).
Except in a few simple cases, it was only proved recently that the necessary ray
expansion represented the complete response (Cisternas, Betarcourt & Leiva 1973;
Kennett 1974), although, of course, its validity had long been assumed. Hron (1972)
has obtained useful formulae for enumerating the rays in a ray expansion. Vered &
Ben-Menahein (1974) have used these formulae with the generalized ray method.
When the generalized ray method is applied to a spherically symmetric inhomogeneous earth model several approximations are necessary or desirable. In this paper
we only consider in detail the approximations used in a plane model. The additional
steps needed for a spherical model have been discussed elsewhere. The normal mode
expansion is converted into a wavenumber integral using the Watson transformation
(Watson 1918). Terms in the ray expansion are defined in terms of their interaction
with interfaces, but in a spherical model two additional effects are important: rays can
20 1
A

202

C. H. Chapman

travel around the Earth passing the same point more than once and rays can travel
through the centre of the Earth being ' reflected ' without interacting with an interface
or velocity gradient. The rainbow expansion (Bremmer 1949) and ray expansion
(Cisternas et al. 1973; Kennett 1974) together split the response into terms which
represent each ray separately. The generalized ray method can only be applied approximately to the spherical model (Gilbert & Helmberger 1972) and in general different
approximations are valid in different regions (Scholte 1956; Duwalo & Jacobs 1959;
KnopofT& Gilbert 1961; Nussenzveig 1965; Phinney & Alexander 1966; Chapman &
Phinney 1972). In order to avoid these difficulties, it is preferable to apply the Earthflattening transformation to the wave equation (Miiller 1971; Chapman 1973; Wiggins
& Helmberger 1974). The problem is now equivalent to a plane model except that the
Legendre function must be approximated by the Bessel function (Szego 1934), and the
solution of the wave equation is only approximate (Chapman 1973). In both
cases corrections can be written as a series in inverse powers of s, the Laplace time
variable, and can be included using the same method as the plane-wave approximation.
In this paper we only discuss wave propagation in a fluid. The same techniques
can be applied to P,SV and SH waves and the equations have been given in Paper I.
The converted waves in the P-SV problem present an additional difficulty but do not
contribute to the first motion. First we discuss the source wave from a pressure source
in a homogeneous fluid. Apart from the complete generalized ray solution, we also
discuss the plane-wave, first-motion and geometrical approximations. The same
approximations are useful for each term in the iterative solution in an inhomogeneous
medium. We show that the geometrical approximations for the infinite series of
multiple reflections does in fact converge to the geometrical ray amplitude. We find
that the first-motion approximation gives a result which is more generally valid than
geometrical ray theory but is almost as simple to evaluate. Numerical results are
given for the simple model considered in Paper I together with some results from a
layered model for comparison.
The notation in this paper follows closely that in Paper I except for certain important
exceptions discussed in Appendix A. This paper contains some of the results on ' The
long-period approximations in body wave theory ' referenced in Paper I.

2. The direct ray

We consider a perfectly elastic, isotropic fluid in which the wave velocity and
density are only functions of the vertical co-ordinate, i.e. a(z) and p(z). An infinitesimal
pressure source of volume Vois located on the z-axis at z = zo. The problem is axially
symmetric and therefore the finite Fourier transform with respect to the azimuth, 4,
and the inverse Fourier sum (A2) are trivial and will be omitted. Taking the Laplace
transform with respect to time, t , (Al), and the modified Laplace-Bessel transform
with respect to the cylindrical radius, x , (A3), we obtain from the equations of motion
and the constitutive equations the well-known ordinary differential system (Gilbert &
Backus 1966)
dv
- = SAV+W.
dz

In this equation, the vector vT = (su,, a)* contains the transformed vertical velocity,
ti,, and the pressure, -n. We also have
WT =

(- (- + $),
1

dfx

p s dx

In Paper I equation (Al) is in error

-jz)

203

Generalized ray theory

where f, and fz are the transformed body-force components and p is the density. The
matrix A is

where 4.2 = u2- p 2 . Here a(z) = ct-l(z) is the wave slowness, and p and qa are the
horizontal and vertical components of the wave slowness. The transformed horizontal
component of displacement is obtained from the pressure using

except at the source.


Introducing an infinitesimalsource of volume Vo on thez-axis at z = zo, with a pressure
Po(t), we find

The prime ' in equations (4) and (19) and Appendices B and D represents
differentiation with respect to z. In order to discuss solutions of (1) in inhomogeneous
media we will first outline the solution in a homogeneous medium. In equations
(5)-( 15) c1 and p are independent of z.
The matrix A has eigenvalues f q a (where the physical Riemann sheet is defined
with Re(sq,) 2 0) such that
AN = Nq
(5)
where q is the diagonal matrix of the eigenvalues

and N has the eigenvectors as columns

The normalization of N will be justified below.


differential system (1) (Gilbert & Backus 1966) is

A fundamental matrix of the

F(z) = N exp [sq(z-z,)]


(8)
where it is convenient to measure the phase relative toz,. Using the standard solution
of the inhomogeneous equation (l),

(zl is arbitary) and applying the radiation condition as z + fco we obtain

204

C. H. Chapman

The discontinuity at z = zo is caused by the source (4) and the fundamental matrix
is given by (8). Applying the inverse transformation (A3b) to (lo), using (3) for the
horizontal component, we obtain for the Laplace transform of the displacement at
(x, 2)

where the sign of the u, component depends on whether z 2 zo. The vector u has
components u, and u,. Methods of solving the inverse transformations (11) and (Al)
which can be extended to problems including reflected and transmitted rays have been
given by a number of authors (Cagniard 1939, 1962; Pekeris 1955a, b; de Hoop 1960).
We follow most closely the notation of Helmberger (1968) and will call the technique
the generalized ray method. First we note that the factor s2 Po(s) is equivalent in the
time domain to convolution with Po(t) and double differentiation which we shall
write D 2 Po(r)*. The inverse Laplace transform of the modified Bessel functions are
known (Erdtlyi er al. 1954, p. 278, equation (18)) and we obtain

where 9 = px+q, Iz-zol. Prior to the inverse transformation (Al), the p-contour
must be distorted from the imaginary axis to the Cagniard contour C defined by
Im(9) = 0. The integrals in (12) can be evaluated by a change of variable (Pekeris
1955a, b) and we obtain

where R is the length of the ray (R2 = x2+ (z-z0)) and i its angle with the vertical
( i = sin-(x/R)).

3. The plane-wave and first-motion and geometrical approximations


In general when we have a more complicated integral than (1l), three approximations are useful: the plane-wave, the first-motion and the geometrical approximations.
The plane-wave approximation can be obtained using the asymptotic expansion
for the modified Bessel functions in (1 1) (Abramowitz & Stegun 1965, equation (9.7.2))
or equivalently expanding the integrand of (12) assuming It-91 < 12~x1(Helmberger
1968). In the first case we obtain a series in inverse powers of s and in the second case
the integral (12) reduces to a series of convolution integrals. In general, the former
method is to be preferred as it can be used for any terms in the response which can
be expressed in inverse powers of s although, of course, the methods are essentially
the same. Examples of terms which can be expressed as inverse powers of s are the
WKBJ expansion (Richards 1971a), the spherical wave functions (Gilbert &
Helrnberger 1972), the Earth-flattening approximation (Chapman 1973), and corrections due to gravity (Gilbert 1967). Applied t o (11) or (12) we obtain

Generalized ray theory

* -(Im[--(
1
(2xP

)] +-Im[--(
D-'

P+ dP

8~

qa d9 1 4 .

1 dP 3P
P'qad9 f q a

)I+...).

205

(14)

The advantage of an expression such as (14) over (12) is that the complex integral is
avoided (all the time convolutions can be combined into one step), and it has been
widely used by Wiggins & Helmberger (1974). A disadvantage is that we have used an
asymptotic expansion and ultimately it will diverge. However, the magnitude of the
next term can always be checked (although from the literature it appears this has not
been done) and usually the first term is sufficient.
The second approximation, the first-motion approximation comes from expanding terms in (14) near the arrival time, t = aR (Knopoff & Gilbert 1959). Then
u(2) =

where P , ( t ) = DP,(t).
Finally, the geometrical approximation, is the solution obtained assuming
geometrical ray theory. Here, it is identical to (1 5) but we shall see that for a turning
ray they differ.
The solution for the direct ray in a homogeneous fluid is, of course, trivial. We have
outlined it in some detail as the notation and methods will now be extended to the
more general problem.

4. The iterative solution


Various methods of solving the differential system (1) are known when A is a function
of z. In this paper we shall ignore the possibility that A is discontinuous, i.e. interfaces
exist, as the ray expansion has been well studied elsewhere. Cisternas et al. (1973) and
Kennett (1974) have derived the ray expansion, Hron (1972) has indicated how these
rays may be enumerated, and Vered 8z Ben-Menahem (1974) have applied these
results to the generalized ray method. We shall assume that A is a continuous function
of z. The technique described in this paper could be applied to each ray in a ray
expansion.
In general, differential equations such as (1) can be solved iteratively using Picard's
method of successive approximations (Coddington & Levinson 1955, pp. 11-12).
Applied directly to (1) this will not be particularly useful as the fundamental matrix,
such as (8), will be developed as a power series expansion in s. We use (8), however,
to suggest how the differential equation (1) can be converted into a form for which
Picard's solution converges rapidly. Suppose we write
where
(17)
20

and the definitions (6) and (7) have been extended to include the z dependence. In a
homogeneous region the components of r(z) are constants except for discontinuities
at the source (10). Substituting (16) in (1) using (5) we obtain
dr _
- Br+s

dz

206

C. H. Chapman

SOURCE

FIG.1. A diagram of the iterative solution (20). The k-th iteration contains signals
reflected k times by the velocity gradient.

where

B = -exp [ -sQ] N- N exp [sQ]


and

s = exp [ - s Q ] N-w.
Picards iterative solution to (18) is identical to the Neumann series for the Volterra
integral equation equivalent to (18) (Morse & Feshbach 1953, p. 905). The iterative
solution for r(z) is given by
z

r(k+

1)(z) = r(o)(z) +

where

J ~ ( 5 r(k)(c)
)
dc

(20)

1
2

r(O)(z) =

s(5) dc

in general, and

for the source (4) and we have assumed the source is located in a homogeneous region
(the subscript 0 refers to values at z = zo). Physically the components of r represent
the amplitudes of the waves travelling in the two z-directions and each iteration adds
another reflection from the velocity gradient (see Fig. 1). Because N has been
normalized so that the energy flux in the z-direction of each column vector is
independent of z (Biot 1957; Alsop 1968) the diagonal elements of B are zero.
Therefore each iteration adds only the next reflections and does not correct the
previous solution for its normalization. We shall call the difference between r(k) and

Generalized ray theory

207

r('-') the k-th order reflections. The off-diagonal elements of B have been given by a
number of authors and interpreted in terms of the reflections from the velocity gradient
(Brekhovskikh 1960, p. 204; Scholte 1962; Chapman 1969; Richards 1970). Scholte
(1962) and recently Richards & Frasier (1976) have obtained solutions in the time
domain for partially reflected signals. However, in those papers it is assumed that the
solution (20) becomes invalid at the turning point of a ray where B is singular. In
Paper I and this paper we show that the iterative solution remains a useful approximation provided the singularity is properly evaluated.

5. The first-order reflections


The source rays are contained in the zeroth term, do),given by (21). The inversion
of this term is simple as it does not include B and we proceed immediately to the firstorder reflections. We assume, to make our notation simplier, that the medium is
homogeneous above the source, i.e. B(z) = 0 for z > zo. Thus the first-order
reflections are given by

Substituting this in (16) we obtain

for the first-order reflections,


where

and

The subscriptz refers to the receiver depth. The depth integral in (23) takes the place of
the generalized reflection coefficient in the simple problem of reflection from an
interface. Taking the inverse Laplace transform of (23) we obtain

where

208

C. H. Chapman

FIG.2. A diagram of the depth limits z,,,,, and z., . The ray paths shown have
travel times 9 and satisfy Snells law. The lower limit is reflected by the velocity
gradient and the upper limit has a horizontal head wave .

and p is defined by Im(S(p, x, 0) = 0. In Paper I we derived the plane-wave approximation to (26) as the limit of reflections from thin homogeneous layers. That method
was instructive but clearly not essential. Keeping 5 fixed, equation (27) defines
Cagniard contours with p a function of the real variable 9. For 9 small, the
Cagniard contour is on the real p-axis, the integrand is real and the contribution zero,
i.e. the signal reflected from a depth 5 is causal. Alternatively, keeping 9 fixed,
equation (27) defines isochrons with p a function of the depth 5. In Paper I we have
discussed the properties of an isochron. Outsidea certain range of depths, the isochron
lies on the real p-axis and again the integrand is real and the contribution zero, i.e
reflections from these depths arrive later than the given time 9. In order to simplify the
discussion we consider the case when only one geometrical ray path to the receiver
exists. Then, equation (26) can be rewritten

where trayis the geometrical arrival time, and zmin(9,x ) < 5 < zmaX(9,
x ) is the range of
depths from which reflections arrive at or before 9. The isochron begins and ends on
the real axis for 5 = zminand z,,
and p = pminand pmaxrespectively. These solve
9 = 9(p,,,, x, z,,) with pmax= a(z,,J for the upper limit. This represents a ray
path with a horizontal segment at 5 = z,.
At the lower limit we have a reflection
from the velocity gradient and
9 = Q@min, x, zmin)
and

209

Generalized ray theory

km

Zray

-2000.-

FIG.3. The range f the 6- 5 integral (28). The numerical values are those used n
Fig. 4.

(see Paper I, Appendix A and Fig. 2). The variables pmin,


zmin,
pmax
andz,,, are functions
of x and 9. The range of the double integral is shown diagramatically in Fig. 3. At the
geometrical arrival time
fray

a9
- (pray,
aP

= $(Pray,

X,zray) = 0

X , zray)

with

Pray

= 4zray)

and zraY
is the turning point of the ray. The variables fray,prayand zraY
are only functions
of x. The geometrical range is given by

where zp is the turning point such that p = a(z,) and we see that

6. The plane-wave, first-motion and geometrical approximations


The first term in the plane-wave approximation to (28) is

210

C. H. Chapman

which is equivalent to the result derived in Paper I (equation (40)) except that the
' head wave ' term has not been explicitly written in (31). The next term in the planewave expansion is simply obtained as in (14). In Paper I we derived (31) as the limit
of reflections from thin layers and included a term for the head wave which exists
for the shallow reflections. The integrand of (31) has a singularity at its upper limit
and if the isochron loops this pole and ends on the real p-axis, the contribution is
equal to the head-wave limit. Thus in (28) and (31) the upper limit is infinitesimally
above zmaxso the pole is included, whereas in (40) of Paper I the pole is not included in
the integral. In Appendix B details of the integral (31) and its numerical evaluation are
discussed. In Appendix C details of the additional integral in (28) are given.
Using the method outlined in Appendix B we can obtain the first-motion approximation to (31) which is
U(t) =

VO

4 2 U o 2 ( P z Po)*

1
tf

D' Po(t) * -

with p = praywhere not otherwise indicated. In the final terms, p = aminor amaxare
solutions of t = 9(p, x,z,) (see Appendix B).
As a further approximation we can substitute estimates for aminand amax(B8).
Then (32) reduces to

Apart from the factor 4 3 this is the geometrical ray result for the amplitude in an
inhomogeneous medium with the source ray (15). We shall call (33) the geometrical
approximation for the first-order reflections. The geometrical spreading factor
ldX/dpl is obtained from (29). This result was derived in Paper I by a less
elegant method. The factor of 7c/3 exists because we have only included the first-order
reflections (22) in modelling the response of the inhomogeneous medium. We now
investigate the next term.

7. The third-order reflections


As we have assumed that the medium is homogeneous above the source and we
will take the receiver above the source, i.e. z 2 zo, we need not consider the second
term, and the next significant term is the third-order reflections. In general only the
odd-order reflections contribute to the first motion after tray. The plane-wave
approximation for the third-order reflections is

211

Generalized ray theory

and p is defined by Im (Q3) = 0. Within the volume V, p is complex whereas outside V,


p is real and the contribution to the integral would be zero.
The details of the volume V , the first-motion approximation and the numerical
evaluation of (34) are sufficiently intricate that they are relegated to Appendix D.
Writing
u(t) =

VO

4n2 cfo2(Pz PO)*

o2Po@) * f+

we note that the first-motion approximation for the first-order reflections (32) is n/3
times this. The first-motion approximation for the third-order reflections (D1 d) is
-n3/648 times (36). Generally, using the same method as in Appendices B and D
we obtain the result that the first-motion approximation for the (2E+l)-th order
reflections is
c21+1

=2

(-1)l

(,)2,+1

(21+1)!

(37)

times expression (36).


Recognizing that
.
C
c21+1= 2 sin
1=0
00

n
-

we deduce that equation (36) is the solution for the first motion of the sum of all
multiple reflections, i.e. the solution of equation (18). The geometrical approximation
gives (using (B8))

which is just the result obtained from geometrical ray theory. The first-motion
approximation (36) can be obtained by three methods: Chapman (1975) derived it using
the generalized ray method as in this paper, Wiggins (1975, 1976) deduced the same
result using a wave-front argument which he calls ' quantized ray theory ' or ' disk ray
theory ', and in Appendix E we show that an identical result can be obtained from the
WKBJ approximation.
The first-motion approximation (36) is a very useful result. When the geometrical
approximation (38) is valid, the two approximations are equivalent. But it is well
known that the geometrical approximation is invalid at caustics, shadows and critical
points (in general at discontinuities in the travel-time curve). At these points, the
first-motion approximation (36) remains useful. Illustrations of this will be published
elsewhere. It should be noted that (36) can be evaluated using the travel-time curve
and that the only complication compared with (38) is one convolution integral.

212

C. H. Chapman
I

uz

114

1.15

1.18

1.17

1.16

1.19

I.

ux
uz
1220.

1230.

1240.

1250.

1260.

FIG.4. Theoretical seismograms for a homogeneous spherical model. The range is


0 = 70, velocity a. = 6 km s-l and the radius ro = 6371 km. The exact solution
is shown with a fine line in (a), (c), (d) and (e) and in (a) and (c) the first-order
geometrical approximation is also shown. The lower and left scales follow equation
(40),and the upper and right scales are in dimensionless units for a unit sphere of
unit velocity. (a) Exact first-order seismograms (28) for u, and uz. The source is
discussed in the text (39). (b) Second term in the plane-wave expansion for the
first-order seismograms ( x 100). (c) First-order, plane-wave approximate seismograms for a plane layered model with d = 100 km (see Paper I for details).
(d) Plane-wave, first- and third-order seismograms (31) and (34). (e) As (d) but
for homogeneous layers d = 100 km. (f)The exact solution (40) and the firstmotion approximation (36) (see also (41)).

Generalized ray theory

213

8. Numerical results
All the numerical results in this paper are for an idealized but theoretical useful
example. We consider, as did Helmberger (1973) and Paper I, the direct ray in a
homogeneous medium. The solution has already been found (13). An alternative
approach is to suppose the source is located at r = ro in a spherical model and use the
techniques employed in a spherical Earth model (Muller 1971; Chapman 1973;
Wiggins & Helmberger 1974). After the Earth-flattening transformation the plane
model is inhomogeneous and the solution can be written as (28). Further details
concerning the Earth-flattening transformation are given in Appendix F.
The source strength is chosen to give a unit displacement step function at unit
distance, i.e.

where c is unit area. Thus the exact solution (13) reduces to


t H (t-LZR)

and numerically 1 1 = R-' at = aR.


The methods used to evaluate the response integrals numerically are discussed in
Appendices B, C and D. The spherically symmetric, homogeneous model has velocity
cl0 = 6-0km s - l with the source and receiver at radius ro = 6371 km separated by
8 = 70" (the numerical values used in Helmberger 1973, and Paper I). In Figs 4 and
14 we have also included axes for the dimensionless solution where the sphere has
unit radius and velocity. The calculations are performed after the Earth-flattening
transformation (Appendix F).
In Fig. 4(a) we have evaluated the exact first-order reflection result (28). The
difference between this and the plane-wave approximation (31) would not be visible
in this graph, but the second term in the plane-wave expansion is shown in Fig. 4(b).
Note that because of the integral operation (D-' as in (14)) this is a ramp rather than a
step. In Fig. 4(c) we have included for comparison, seismograms calculated for the
same model approximated by homogeneous layers with a thickness 100 km. The factor
of 7c/3 N 1.047 in the first motion of Fig. 4(a) (32) compared with the exact result (40)
is obvious. When we include the third-order reflections (34) we obtain the result in
Fig. 4(d). The first motion is now n/3-1r3/648 N 0.99935 times expression (40),
and the error has been reduced from about 5 per cent to 0.06 per cent. However, at
longer periods the error is still significant. It would appear, at least to the author, that
numerical evaluation of higher-order reflection integrals is not practical on a regular
basis. The generalized ray method would be extremely expensive if multiple reflections
had to be included to obtain the required accuracy. We have not attempted to quantify
these errors as they depend on the model. This paper is concerned with the firstmotion approximations which are not model dependent. In Fig. 4(e) we have included
results for the third-order reflections in an approximate model with layer thickness
100 km. Despite this coarse layering, over 1000 generalized rays were included and the
computation times for Figs. 4(d) and (e) were comparable. Although the limiting
behaviour for fine layers is not obvious from Fig. 4(e), it is consistent with Fig. 4(d).
Finally in Fig. 4(f) the first-motion (36) and exact solution (40) are shown. The
geometrical approximation (38) would be a simple step function. Using the velocity

214

C. H.Chapman

model (F3), an expansion for the first-motion solution (36) can be obtained. Then
u ( t ) = z ( S i n i ) H ( t - a R ) [1+-(t-uR)
U

cos i

=!(sin
- i,
R cos i
N

R
H(t-uR)

[1- R ($+& tan i)(r-uR) 1


U
-

5
R (Sini)H(t-aR)
cos i

Here, expression (41a) is the exact solution (40), (41b) is the first-motion approximation (36) and (41c) is the geometrical approximation (38). In this example, the
geometrical approximation is better than the first-motion approximation but this
is not generally true. If the range (29) is a discontinuous function of ray parameter (p),
or if its gradient (the spreading factor) is zero or infinite, then the first-motion approximation is better (Chapman 1976). The agreement between the various approximations
is very good here only because the model is so simple.

Conclusions

In this paper, the generalized ray method has been developed for a vertically
inhomogeneous media without interfaces. The ray expansion and Earth-flattening
transformation needed to apply the method to a realistic Earth model have been
studied elsewhere. It is straightforward to add them to the theory presented in this
paper.
The exact response can be obtained using the generalized ray method as an infinite
series of multiple integrals representing multiply reflected signals. In practice only a
very limited number of terms can be evaluated numerically although various approximations make the problem more tractable.
The plane-wave expansion permits a complex wavenumber integral to be replaced
by an asymptotic series of convolution integrals. Normally the first term is sufficient
and the convergence can always be checked by the next term. A considerable saving
in computation results as a single convolution replaces a complex integral for every
time in the exact result.
Even after the plane-wave expansion we are still left with a series of multiple depth
integrals. Only the first- and third-order reflections have been evaluated numerically.
However, a very useful first-motion approximation is possible which provides an
inexpensive method of computing synthetic seismograms and proves the convergence
of the infinite series, at least for the first-motion. The first-motion approximation
for multiple reflections converges as the terms in the Maclaurin series of 2 sin(n/6),
The first-motion approxi.e. successive terms are in the ratio - (21+ 1)(21)(6/~)~.
imation, which can be computed directly from functions derived from the
travel-time curve, is equivalent to geometrical ray theory if the latter is valid. However,
it remains useful at caustics and shadows when geometrical ray theory must be
extended beyond the second-order, saddle-point method.
Although the original intent of this paper was to develop generalized ray theory
completely for a vertically inhomogeneous medium, undoubtably the first-motion
approximation will be the most useful result contained in it. This approximation
has now been obtained by other methods (Wiggins 1975, 1976; Chapman 1976 and
Appendix E) and proved to be valid at caustics and shadows. As an approximation, it

Generalized ray theory

215

is obviously most valuable and inexpensive. Its limits of validity require further investigation. A study of the inversion of body-wave data using this approximation
will be published elsewhere.
Acknowledgments

I should like to thank Professor B. A. Bolt who provided computer funds from
National Science Foundation grant GA 4101A1 for computations done at the
Computer Center of the Lawrence Berkeley Laboratory where this study was initiated.
Other computations were performed at the Computing Centre of the University of
Alberta and the University of Toronto Computer Centre partly financed by National
Research Council of Canada operating grants, A7475 and A9130 respectively.
References

Abramowitz, M. & Stegun, I. A., 1965. Handbook of mathematical functions, Dover


Publications, New York.
Alsop, L. E., 1968. An orthonormality relation for elastic body waves, Bull. seism.
SOC.Am., 58, 1949-1954.
Bauer, F., 1961. Romberg integration, Algorithm 60, Communs. Ass. Comput.
Mach., 4, 255.
Bessonova, E. N., Fishman, V. M., Johnson, L. R. Shnirman, M. G. & Sitnikova,
G. A., 1976. The tau method for inversion of travel times 11. Earthquake data,
Geophys. J. R . astr. SOC.,in press.
Bessonova, E. N. Fishman, V. M., Tyaboyi, V. Z. & Sitnikova, G. A., 1974, The tau
method for inversion of travel times I. Deep seismic sounding data, Geophys.
J. R . astr. SOC.,36, 377-398.
Biot, M. A., 1957. General theorems on the equivalence of group velocity and energy
transport, Phys. Rev., 105, 1129-1 137.
Brekhovskikh, L. M., 1960. Waves in layered media, Academic Press, New York.
Bremmer, H., 1949. Terrestrial radio waves, Elsevier, Amsterdam.
Budden, K. G., 1966. Radio waves in the ionosphere, p. 325, Cambridge University
Press.
Bullen, K. E., 1963. An introduction to the theory of seismology, Cambridge
University Press.
Cagniard, L., 1939. Rdjlexion et rdfraction des ondes sdismiquesprogressives, GauthierVillars, Paris.
Cagniard, L., 1962. Reflexion and refraction of progressive seismic waves, trans.
E. A. Flinn & C. H. Dix, McGraw-Hill, New York.
Chapman, C. H., 1969. Seismic wave drflraction theory, PhD thesis, Cambridge
University.
Chapman, C. H., 1973. The Earth-flattening transformation in body wave theory,
Geophys. J. R. astr. SOC.,35, 55-70.
Chapman, C. H., 1974a. Generalized ray theory for an inhomogeneous medium,
Geophys. J. R . astr. SOC.,36, 673-704.
Chapman, C. H., 1974b. The turning point of elastodynamic waves, Geophys. J. R.
astr. SOC.,39, 613-621.
Chapman, C. H., 1975. Greens functions for the impulsive wave equation using the
asymptotic expansion or the ray series, Bull. Can. Ass. Phys., 31, 47.
Chapman, C. H., 1976. A first motion alterative to geometrical theory, Geophys.
Res. Lett., 3, 153-156.
Chapman, C. H. & Phinney, R. A., 1972. Diffracted seismic signals and their
numerical solution, Methods in Computational Physics, 12, 165-230, Academic
Press, New York.

216

C. H.Chapman

Chester, C., Friedman, B. & Ursell, F., 1957. An extension of the method of steepest
descent, Proc. Camb. phil. Soc., 53, 599-61 1 .
Cisternas, A., Betarcourt, C. & Leiva, A., 1973. Body waves in a ' real Earth ' Part I,
Bull. seism. SOC.Am., 63, 145-156.
Coddington, E. A. & Levinson, N., 1955. Theory of ordinary differential equations
pp. 11-12, McGraw-Hill, New York.
Duwalo, G. & Jacobs, J. A., 1959. Effects of a liquid core on the propagation of
seismic waves, Can. J. Phys., 37, 109-128.
ErdClyi, A., Magnus, W., Oberhettinger, F. & Tricomi, F. G., 1954. Tables of integral
transforms, Vol. 1, McGraw-Hill, New York.
Gilbert, F., 1967. Gravitationally perturbed elastic waves, Bull. seism. SOC.Am., 57,
783-794.
Gilbert, F. & Backus, G. E., 1966. Propagator matrices in elastic waves and vibration
problems, Geophysics, 31, 326-332.
Gilbert, F. & Helmberger, D. V., 1972. Generalized ray theory for a layered sphere,
Geophys. J. R . astr. SOC.,
27, 57-80.
Helmberger, D. V., 1968. The crust-mantle transition in the Bering Sea, Bull. seism.
SOC.Am., 58, 179-214.
Helmberger, D. V., 1973. Numerical seismograms of long period body waves from
seventeen to forty degrees, Bull. seism. SOC.Am., 63, 633-646.
Hoop, A. T. de, 1960. A modification of Cagniard's method for solving seismic
pulse problems, Appl. scient. Res., B8, 349-356.
Hron, F., 1972. Numerical methods of ray generation in multilayered media, Methods
in Computational Physics, 12, 1-34, Academic Press, New York.
Jeffreys, H., 1939. The times of core waves, Mon. Not. R. aftr. SOC.,Geophys. Suppl.,
4, 548-561.
Johnson, L. E. & Gilbert, F., 1972. Inversion and inference for teleseismic ray data,
Methods in Computational Physics, 12, 231-266, Academic Press, New York.
Kennett, B. L. N., 1974. Reflections, rays and reverberations, Bull. seism. SOC.Am.,
64, 1685-1696.
Knopoff, F. & Gilbert, F., 1959. First motion methods of theoretical seismology,
J. acoust. SOC.
Am., 31, 1161-1168.
Knopoff, L. & Gilbert, F., 1961. Diffraction of elastic waves by the core of the Earth,
Bull. seism. SOC.Am., 51, 35-49.
Langer, R. E., 1937. On connection formulae and the solution of the wave equation,
Phys. Rev., 51, 669-676.
Morse, P. M. & Feshbach, H., 1953. Methods of theoreticalphysics, p. 905, McGrawHill, New York.
Muller, G., 1970. Exact ray theory and its application to the reflection of elastic
waves from vertically inhomogeneous media, Geophys. J. R . astr. SOC.,21,
26 1-283.
Muller, G., 1971. Approximate treatment of elastic body waves in media with spherical
symmetry, Geophys. J. R . astr. Soc., 23, 435-449.
Nussenzveig, H. M., 1965. High frequency scattering by an impenetrable sphere,
Ann. Phys., 34, 23-95.
Pekeris, C. L., 1955a. The seismic surface pulse, Proc. Nut. acad. Sci. USA, 41,
469-480.
Pekeris, C. L., 1955b. The seismic buried pulse, Proc. Nat. acad. Sci. USA, 41,
629-639.
Phinney, R. A. & Alexander, S . S., 1966. P wave diffraction theory and the structure
of the core-mantle boundary, J. geophys. Res., 71, 5959-5975.
Phinney, R. A. & Cathles, L. M., 1969. Diffraction of P by the core: a study of long
period amplitudes near the edge of the shadow, J. geophys. Res., 74, 1556-1574.

217

Generalized ray theory

Richards, P. G., 1970. A contribution to the theory of high frequency elastic waves,
PhD thesis, California Institute of Technology.
Richards, P. G., 1971a. Waves in stratified media, Geophysics, 36, 798-809.
Richards, P. G., 1971b. An elasticity theorem for heterogeneous media, with an
example of body wave dispersion in the Earth, Geophys. J. R. astr. Soc.,22,453-472.
Richards, P. G., 1973. Calculation of body waves, for caustics and tunnelling in core
phases, Geophys. J. R. astr. SOC.,35, 243-264.
Richards, P. G., 1974. Weakly coupled potentials for high-frequency elastic waves
in continuously stratified media, Bull. seism. SOC.Am., 64, 1575-1588.
Richzrds, P. G. & Frasier, C. W., 1976. Scattering of elastic waves from depthdependent inhomogeneities, Geophysics, in press.
Scholte, J. G. J., 1956. On seismic waves in a spherical Earth, Meded. Verh. K. ned.
met. Inst., 65.
Scholte, J. G. J., 1962. Oblique propagation of waves in inhomogeneous media,
Geophys. J. R . astr. SOC.,
7,244-261.
Spencer, T., 1960. The method of generalized reflection and transmission coefficients,
Geophysics, 25, 625-641.
Szego, Von G., 1934. Uber einige asymptotische Entwickhungen des Legendreschen
Funktionem, Proc. Lond. math. SOC.,2, 427-450.
Vered, M. & Ben-Menahem, A., 1974. Application of synthetic seismograms to the
study of low-magnitude earthquakes and crustal structure in the northern RedAm., 64, 1221-1237.
Sea region, Bull. seism. SOC.
Wasow, W., 1965. Asymptotic expansions for ordinary differential equations, Pure
and applied mathematics, Vol. 14, Interscience, Wiley, New York.
Watson, G. N., 1918. The diffraction of electric waves by the Earth, Proc. R. SOC.
London, A95, 83-99.
Wiggins, R. A., 1975. Body-wave synthetic seismogram calculations by quantized
ray theory, Bull. Can. Ass. Phys., 31, 47.
Wiggins, R. A., 1976. Body wave amplitude calculations-11, Geophys. J. R. astr.
SOC.,
46, 1-10.
Wiggins, R. A. & Helmberger, D. V., 1974. Synthetic seismogram computation by
37, 73-90.
expansion in generalized rays, Geophys. J. R. astr. SOC.,

Appendix A
Notation
Several important changes in notation have been made in this paper compared
with Paper I. These have been made to simplify expressions or to make the meaning
more logical.
The Laplace transform with respect to time and its inverse, are defined as
OD

u(s) =

u(t) exp ( - s t ) dt

u(r) =

271i

iw
.

u(s) exp (st)ds.

MI)

- ioo

In common with other branches of theoretical physics we have found it sufficient to


distinguish a function from its transform by its argument and/or context. Therefore
tildes have not been used.
The finite Fourier transform with respect to the azimuthal angle, 4, and the
inverse series, are
2%

's

u(l) = - u(4) exp (--W)@ u(4) =


u(Z) exp (il4)
(A2)
l=-m
2n 0
where the position of the factor (2n) is chosen such that u(0) = u(4) if u is independent
of 4 (as in the problem discussed in this paper).

218

C. H. Chapman

We have found it inconvenient to use r for the cylindrical radius (reserving this for
the spherical radius), or p (used for density), and have used, as Wiggins & Helmberger
(1974) did, x . Then in either Cartesian or cylindrical geometry x is the horizontal
range. We define the modified Laplace-Bessel transform with respect to the
cylindrical radius x as
m

and the inverse


-im

Generally in the seismology literature, ci and 8 are used for the P and S wave
velocities. We have followed this notation but have also found it convenient to use
a and b for the P and S wave slownesses (Phinney & Richards, private communication
and Russian literature), i.e.

Many equations are algebraically simplified if a and b are used.


The use of p for the horizontal wave slowness (ray parameter) is very common,
but the author has previously used q and A for the vertical slowness and matrix, e.g.
Paper I. In this paper we have used q and q for the vertical slowness and matrix (6).
This suggests using Q for the slowness integral (17) which is consistent with the use
in Richards (I 97 1a) and Chapman (1973).
Although Helmberger (1968) used z for the phase functions such as (27), we have
used 9 (Chapman 1976) to avoid confusion with the tau function (B20) which is now
widely used in the inversion of travel-time data (Johnson & Gilbert 1972; Bessonova
et al. 1974, 1976).
In Paper I, the limits of the depth integral were defined as zminand z, (Figs 7, 8
and 9 in Paper I) for the minimum and maximum depths, respectively. The choice was
illogical as the z - co-ordinate was not depth but was positive upwards and it resulted in
pmin> pmax(Fig. 13 in Paper I) for the corresponding ray parameters. Therefore in
this paper we have interchanged the meaning of the subscripts (28).
In Paper I we use the symbol M in the differential system (I) but here we replace
it by A (2) (as in, for instance, Gilbert & Backus 1966) (retaining M for use as the
moment tensor).

Appendix B
The first-order depth integral

In order to evaluate the integral (28) or (31) for the first-order reflections we must
investigate the properties of the first-order isochron. It is defined by Im (9(p, x, c)) = 0
where 9 is defined in (27). In general when we evaluate the integral (28) or (31) this
must be solved for complex p as a function of time and depth. Numerically this is
achieved using the Newton-Raphson method with an initial estimate from the first
derivative

219

Generalized ray theory

The end points (pmin,


zmin)and (pmax,
z,,,) defined above are found first, again using the
Newton-Raphson method where in the case of (Pmin,zmiJ both p and 5 must be
iterated. At the end points, the derivative (BI) is infinite ([ = zmln)or zero (5 = zmaJ
and is not useful numerically. Expansions given in Paper I (remembering we have
reversed the meaning of the subscripts ' max ' and ' min '-Appendix A) can be used
to estimate the first complex point, i.e.

where amax= pmax


and atmax
are the slowness and its gradient at = z,,.
In order to
investigate the shape of the isochron, it is convenient to use the first-motion
approximation found in Paper I (equation (85)). This is

Letting

W =

2' 3/6 u,,y

v-plane

FIG.5. The curve Im (u) = 0 defined by equation (BS).

220

10-

C. H. Chapman

1l

p -plane
s/ km

C I
\

I
I

Pmin

,1362

,1364

I
1

Pray

,1366

.I368

pmox

.I370

FIG.6. The first-order isochron 8 ( p , x , 1) = t with t = t,.,+0~01 s and the other


parameters as in Fig. 4. The units ofp are s km- and of 5 are km. For this time
z,.
= -1289.02 km and zmax=-1252.89 km. The ticks mark the depths
z = - 1255 to - 1285 km at 5-km intervals.

equation (B4) reduces to the dimensionless form

with the first-motion condition that w < 1. Ideally we require a solution u(u) but the
basic properties are easily found. The line Im (u) = 0 is the isochron and it meets the
real axis at v = 1 when u = w-, and v N - 1 - 3 w 3 when duldv = 0. The point
u = 0 corresponds to v N w exp ( i 4 3 ) . The complete curve is shown diagrammatically
in Fig. 5, and from a numerical solution for p with 9 = r,,,+O.Ols and the other
parameters as in Fig. 4, in Fig. 6 . We can rewrite (B5) as

v = v - u ~ u = w(I -u)/~ exp (in/3)

(B6)

where

We note that always 161 < 1 and when v is small, V N w exp (in/3). This is confirmed
numerically in Fig. 7, where we have plotted the variable

P =P-am

037)

corresponding to V. We note that for most of the depth range Arg(p) is approximately
constant corresponding to the solution V N w exp (in/3)above. In order to retain
numerical accuracy it is preferable to use jj rather than p as the variable. We shall see
that this is also convenient analytically.
In order to find the end points we use

to estimate z,,, and zmin.In order to estimate pmin(< amin)we integrate the singularity of d2 $/dp2 (as in equation (51) Paper I) and obtain

22 1

Generalized ray theory

FIG.7. The functionfl defined in (B7) correspondingto Fig. 6.

and
s ( p m i n , X, Z m i J N $(amin,

Fmin

as

aP

x, zmin)+ __ -

(amin,

x, Zmin)

(B9b)

We have already noted in Paper I that the depth integral, (28) or (31), has an
integrable singularity at each end point. The important terms are (Paper I)

near zmin,and

near zmax.With changes of variable y = (C-zmin)*for (BlO) and y = (zma,-O* for


(B1l), both singularities can be integrated (Paper I). However, even as a function of y,
the integrand of (28) or (31) varies rapidly and we have found another variable better.

222

C. H. Chapman

2.

I.

RG.8. The fbction P = log,@ correspondingto Fig. 6. The behaviour at infinity


is shown diagrammatically. P, is for E = 0.01 km.

We consider the integral


(B12a)
The factors in the integrands of (28) and (31) which are contained in the function C([)
are slowly varying. They cause no trouble in the numerical integration and can be
treated as constant in the first-motion approximation. We change the variable to j?
defined in (B7) and obtain
(B12b)
where

and

The function r A ( p , 4') is dimensionless and slowly varying. In the first-motion


approximation it reduces to 1/4. Letting P = log,? we obtain
(B12c)
where J ( p , x, [) is the most significant term in the integrand. In Fig. 8 we have plotted
P corresponding to p in Fig. 7. Note that p,,, = 0, so P,,, is infinite. Outside the
range zmin< < z,,
is negative real and, as already noted, the integrand has zero
imaginary part. For [ = z,,,-E (where E is small) p lies just above the positive real
axis, and for 1: =,,,z + E it is negative real. The contour runs in an infinitesimal loop
above the pole at p = 0. Thus for [ 5 zman,
P 4 (- coy0) but including the. pole we
take P,,, = (- co, n) (shown diagrammatically in Fig. 8). The segment from 0 to in

Generalized ray thcory

223

FIG. 9. The integrand of (B12c) corresponding to Fig. 6 for the component uz.
The normalization is such that the geometrical approximation for the integral is
51
cos i
uz = ImJf(P)dP N - . -(t-tray)-* 2: 0 . 8 2 10-3
~
3
R

at infinity just gives the ' head-wave ' term found less elegantly in Paper I (equation
(40)), i.e. from (B12c).
(B15)

Note that Im(Pmin)= x but Im(P) = 7c/3 for most of the depth range (Fig. 8). The
important term in the integrand, J , has the end-point values
(B16a)

(B16b)
and the integrand of (B12c) for integral (31) is shown in Fig. 9. The integrand only
varies rapidly between the end-point values (B16) in the middle of the range and this
is just where P varies very slowly (Fig. 8). Plotting the real part of the integrand
versus imaginary part of P in Fig. 10, this can be seen more clearly.
We have found that despite the integral (B12c) being complex, it can be performed
numerically as efficiently as the real integral. We use the Romberg extrapolation
method (Bauer 1961) where the subdivisions are chosen uniformly in the real variable,
y. Although the steps are not uniform in P , the Romberg algorithm still applies. The

224

C. H. Chapman

FIG.10.The real part of Fig. 9 cs the imaginary part of Fig. 8.

estimated error can be expressed in terms of the step-size in y, and using the trapezoidal
rule, satisfies the necessary condition for the Romberg algorithm that it is second order.
Obviously, however, the numerical integral cannot be taken to the infinite limit H,,,
but must be terminated at P, corresponding to 5 = z,~,-E.
Various methods of
evaluating the end-point integral are possible, e.g. using (B1I), but we have found an
- E < 5 < z,,,, I,.
approach using (B12b) most satisfactory. Call the integral for zmaX
Expanding the integrand we obtain

A and B are real


Although ds/ap and qa have a term O@), these cancel in J ( p , x , i).
constants found numerically from the integrand at p = p,. Evaluating (B17a) we
obtain

As E + 0, we obtain I, + Ipole(B15). The next term is of order E* and the last two
terms e3I2.
The behaviour of the integral shown in Fig. 10, immediately suggests how the first
motion approximation can be obtained. The constant values (B16) are assumed

225

Generalized ray theory

between the end points and Im(F) = n/3 and we obtain

- c(zray)

,
1
[ X-X(Prnax)
+ (-gamin
a

min

(B12d)

Fmin)+

The first term comes from (B16b) with rA(p,,,, z,,,) = 4. At the lower limit we
have used (B16a). We note, however, that other choices of variable F are possible
without altering the basic properties of the integral (B12c). In fact analytically
P = log, ( c q J might be preferred as the argument of the logarithm is then dimensionless but our choice is simpler numerically. The exact form of the second term in
(B12d) differs depending on the choice of P although they are identical to lowest
order in pmin(as written). Using (B9a), the second term in (B12d) can be replaced
x,zmin),to the same accuracy, and by (30) we have
by -I?p/I?3(umin,
(B12e)
This is used in the first-motion approximation (32). Using (B8) we obtain the
geometrical approximation (33). To the same accuracy as (B12e) we can use
Q(uminr
x, zmin)rather than 9(pmin,
x, zmin)(B9b). We note that in general we can
write
(B18)
9(P, x,z,) = P ( x - x ( p ) ) + T(P) = P X + M
where p = u(z,), X t p ) is the geometrical range defined in (29), T(p)is the geometrical
travel time

and

t(p)

is

ZP

=P

(Bullen 1963). Then in equations (B12e) and (32), p = aminor amaxsolve


t = 9(p, x, z,) and are functions of t and x.
The evaluation of the first-motion approximation, (32) and (BlZe), is trivial and
can be derived directly from the travel-time curve. Chapman (1975) and Wiggins
(1975) have obtained this expression independently, the former as in this paper by a
wave-equation approach (when higher-order reflections are included-see Appendix
D) and the latter by a wave-front approach. In Appendix E we obtain the same
result using the WKBJ approximation.

Appendix C
The time integral
The $-integral in equation (28) has integrable singularities at both limits: at the
upper limit 9 = t due to ( t - 9)-* and at the lower limit 9 = traydue to the first-motion
approximation ($-tray)-* (as in (B12e) using (B8)). These can be removed
simultaneously with the variable change

$ = sin-'

29- t - fray
t - tray

so d9 = ((t-$)(9-tray))~d$. The range of the $-integral is -n/2

< $ < n/2.

226

C. H. Chapman

Appendix D

The third-order integrals


The numerical evaluation of (34), or the derivation of an approximate solution,
presents considerable problems. The boundaries of I/ are not plane and the integrand
has singularities due to
Y ~ ( Pti),
, i = L2,3

and
aplaS3.
The integrand is symmetric in and and we restrict the solution to > and
introduce a multiplicity factor, 2. Naturally
> (see Fig. 11). When =
equation (35) reduces to (27) and p lies on the first-order isochron (see Appendix B).
As increases, p leaves the first-order isochron and for cz = z,,,(t, x,
meets the
c2 and
real axis. This may correspond to a ray path with ' reflections ' at
(Fig. 11(a)), or a path with ' reflections ' at and and a ' head wave ' at
(see
Fig. ll(b)) (with special cases when
=
Fig. ll(c), or when we have a turning
Fig. 1 I (d)). Thus for the integral part of (34) we have
point at

el

c3

c2

c3

cz

el,

c3

c2 c3

el c3,

c2

cZ c3,

el, c3)
el,

el

el

c3.

c3,

The
As mentioned above, the type of end point at
= zIopdepends on
and
- area is divided into two regions as shown in Fig. 12. In the lower region, A,
the limit is a ' reflection ' and aS3/dp = 0 and in the upper region, B, the limit is a
' head wave ' and p = ~ ( 4 ' ~ ) .The dividing line satisfies both conditions (Fig. 1 1 (d)).
Thus the integrand of (Dla) has a singularity on the surface zIopdue to either

c3

c2

FIG.1 1 . Diagrams of third-order reflections: (a) with ' reflections ' at 5 1 , and 5 3 ;
(b) with ' head wave ' at I,, ' reflections ' at Cz and
(c) wiih ' head wave ' at
5, = Is, ' reflection ' at c2; (d) ' turning point ' at cl, ' reflections ' at Cz and 13.

c3;

227

Generalized ray theory

5,

____)

Zmin

fray

XZ
,

FIG. 12. The range of


in integral (34). In region A the end point is a
reflection , and in region B it is a head wave .

c3)

el

dp/89,, y,@, 5,) or both (and yA(p, when = g3). In order to evaluate the integral
(Dla) we could remove these singularities with real variable changes (as for the
first-order integral in Paper I). However, the [,-integral is still unsuitable for numerical
2
Then the singularities of y A ( p ,
and yA(p,
both lie
evaluation when
=
near p = a(cl) and the integrand varies rapidly. In the limiting case when
the singularity is of higher order and a special method is necessary. We shall not persue
the details here as although it is the most obvious methods of evaluating (Dla)
it is not necessarily the best. An obvious alternative is to obtain a result comparable
with (B12c). Let us define

c3 el.

ai = a(Ci)

pi = p - a i

Pi= logepi
Then equivalent to (Dla) we have

c3)

with i = 1, 2 and 3

c2)

el c3

(D2)

where r,(p,5) is defined in (B14) and V is the complete volume in a-space. Changing
the variables ai to Pi

228

C. H. Chapman

where

thc third-order equivalent of function (B13). The colume V is now a R3hypersurface


in the R6 = C3 space of the three complex variables, Pi. However, while we
could evaluate (B12c) accurately, the same does not apply to (Dlc). This is
a general problem with singular multiple integrals.
When we change the variable of integration to remove singular behaviour from
the integrand, we in effect transfer the singular behaviour to the new variable of
integration. Thus, in Paper I, we used
dy

dY
-dz
dz

to change the variable of integration and dy/dz contains the singularity of the
integrand. The y-integral can be evaluated accurately because we use the exact Ay's
and not Ay N (dyldz) Az. In other words, we know Jdy exactly. Similarly for the
P-integral (B12c). Although a small number of points may not approximate the
complex line P(5) well, we can still evaluate JdP exactly and, because the integrand is
approximately constart, we can obtain the integral (B12c). This is not necessarily
true for a multiple integral.
In order to evaluate the integral (Dla) accurately we must be able to evaluate
JSSd(, dT2di3. Similarly for (Dlc) we must be able to evaluate JJJdPl dPz dF3. For
the integral (Dla) we can evaluate the volume integral reasonably accurately as the
boundaries of V are smooth and finite. A simple triangulation procedure (filling the
space with tetrahedra) is sufficient and only small errors are made at the boundaries.
A similar triangulation of the P-space is possible, but large errors now exist. A small
number of points not only are a poor apprcximation to the hypersurface but, unlike
the single integral, do not allow JJJdH, dP2 dP, to be evaluated. We have transferred
our problem from the singularities of the integrand to the shape of the hypersurface.
The integral (D lc) is useful though in obtaining the first-motion approximation.
The method is a simple extension of that used in Appendix B and can be extended to
the (21+ 1)-th order reflections. The definitions (B13) and (D3) are easily extended for
the (21+ 1)-th integral. At the lower and upper limits, this function always reduces
to (B16a) and (B16b), respectively. The integral contains (21+ 1) functions T,(p, 1)
which are approximately 1/4. Finally, we need the volume in P-space. At each end
point Im(P) varies from n/3 to n in an analogous fashion to the first-order integral.
'.
Allowing this variation in all the P's we have a hypercube of volume ( 2 ~ / 3 ) ~ ' +But
this hypercube contains (21+ 1) ! hypertetrahedra which are all equivalent by symmetry.
Thus

where we have used the same approximations as in (B12e). Remembering that in


the (21+ 1)-th order reflection, I of the reflection coefficientsare negative (minus sign in
(34)), (Dld) reduces to (36) and (37).

Generalized ray theory

229

Let us return now to the numerical evaluation of (Dla). It is evident from the
above argument that we must choose one variable to remove all the singularities. We
define

Consider a line with the yi's constant. Then da = da, = da, = da, and the line is
parallel to the symmetry line a, = a, = a3. For the isochron defined by (35) we must
have dp = djji = dp, = dp3 and

Substituting in (Dlb) we have

with

dP=-.
F1

djj

iiz F 3

From this we obtain

A complete description of the variable and the integrals in (Dlf) would be too
tedious to include here. We shall only include some general comments.
The line defined by y , and y , intersects the surface 4', = ztopin two points. At the
upper limit J 3 > 0 and at the lower limit J 3 < 0. In general between these limits, the
integrand behaves as in (B12c) (Fig. 9) and only varies rapidly in mid-range. At the
upper limit of y , , J 3 = 0. The line J, = 0 is shown in Fig. 12 and lies in the region
where the end point is a ' reflection '. At the lower limit of y , ( = 0 and C2 = C3), the
third-order isochron is identical to the first-order isochron. Thus, in general, when y ,
is near its lower limit, the P-integral has a ' reflection ' at its lower end and a ' head
wave ' at its upper end. Here
becomes infinite. The last part of the integral is
evaluated using the same technique as (B17a) on the integral (Dle). Taking the
upper limits for the j as negative real, the ' head-wave ' poles are included. Near the
y 1 upper limit, both ends of the p-integrals are ' reflections ' and the complete integral
can be evaluated numerically. At the upper limit, the P-integral has zero length but
has a finite limit.
The complex P-integral in (Dlf) was evaluated using the same numerical methods
as for (B12c). The remaining integrals are real and without singularities (although
variable changes are made to improve their behaviour). The y,-integrals between
y , = 0 and the point where J 3 = 0 are evaluated next using the Romberg extrapolation

230

C.H. Chapman

method (Bauer 1961). And finally the y,-integral between y, = 0 and the source is
evaluated using the Romberg method again. Numerical results (Fig. 4(d)) confirm the
first-motion approximation (Dld).

Appendix E
The WKBJ approximation
In order to solve the differential system (1) we have transformed it into the form (18)
for which the iterative method of solution (20) is appropriate. Despite the singularity
of the kernel in the integral equation (20), the response converges (36) and agrees with
the geometrical approximation for a simple turning ray (38).
The geometrical result can be obtained from a purely geometrical argument (Bullen
1963), but it can also be obtained using the WKBJ approximation to the wave equation
and the second-order, saddle-point method (Richards 1973). At high frequencies, the
wave equation can be approximated by Helmholtz equations for the scalar potentials
(Richards 1974). The WKBJ solutions can be connected to the solutions of Stokes
equation near the turning points or, more generally, we can use the Langer uniform
asymptotic expansion (Langer 1937). This can be extended to higher-order differential
equations such as (1) directly without the introduction of scalar potentials (Wasow
1965; Chapman 1974b). Here we shall only need the high-frequency approximation
for the first term in the expansion which can be obtained using either method.
In the inverse time transform (Al) we write s = -io. Using the WKBJ
approximation, the reflection coefficient is given by (Budden 1966, p. 325, with a sign
change)
- i exp {iWQa(P, z,)).
Using the source wave (10) and this reflection coefficient, we obtain

where we have used the asymptotic expansion of the Bessel function. The usual
method of evaluating an integral like (El) is with the second-order, saddle-point
method. For most values of p the integrand is oscillatory and cancels, and the only
significant contribution comes from the saddle points when the phase is stationary.
The stationary phase condition reduces to equation (30) being zero, i.e. p = pray.
Expanding the phase in a Taylor series to second order about p = prayand evaluating
the saddle-point integral, we obtain the geometrical ray result (38). However, we have
already noted that this is often inadequate particularly at discontinuities in the traveltime curve. Then the first-motion approximation (36) is preferred.
Suppose rather than finding the stationary phase condition in (El) we find the
equal phase condition. Thus we find the values of p such that
t = 9(p, x , zp).
(E2)
Several solutions p = n, (say, with m = 1,2, . . .) may exist, and each 7rm is a function
o f t and x (the symbol 7t without a subscript retains its usual meaning). Note that this
is identical to the definition of pmar(when zp = z,,,), and differs only by second-order
small quantities form the definition of pmin(Appendix B). In general, for each saddle

Generalized ray theory

23 1

point there will be a pair of equal-phase points. We now expand the phase in a Taylor
expansion to first order about p = n, and obtain

* - R1e /

(W'

1
m

(4a,4ao)+

( ) exp [io(x-X(n,))(p-n,)]ripncc,

(E3a)

qa,

for the contribution from one equal-phase point. Remember that the integrand is a
function of time, t , through the definition of z,,, (E2). Evaluating the o-integral we
obtain

Hence we obtain

which is in agreement with (36). The external factors are evaluated with p = x, N pray.
The second-order, saddle-point method can be extended to be useful near caustics
(using the third-order saddle-point method: Jeffreys 1939; Chester, Friedman & Ursell
1957) and near shadows (using the Fresnel integral: Phinney & Cathles 1969; Richards
1971b). Chapman (1976) has shown that including all the equal-phase points, the
result (E3c) agrees with these methods at caustics or shadow edges.

Appendix F
The Earth flattening transformation
The numerical calculations in this paper are for a ray in a homogeneous spherical
model. The same example has been used by Helinberger (1973) and Paper I. This
model transforms into an inhomogeneous plane model and the general method of this
paper is applicable (Fig. 13). It is an important example because the solution is known
(13). The additional approximations introduced by the Earth-flattening transformation are not of direct concern in this paper and have been discussed elsewhere
(Chapman 1973). In this appendix we just list the necessary variable changes and give
some numerical examples to illustrate the importance of the density transformation.
In order to maintain the kinematic properties of geometrical ray theory, the
co-ordinates are transformed as (Muller 1971)

r
r0 = exp

(t)

r,Q = x

where ro is some reference radius, and the velocities as

232

C. H. Chapman

HOMOGENEOUS SPHERICAL EARTH

'FLAT'
0

E A R T H APPROXIMATION
2000

4000

6000

8 0 0 0 krn

-2000

FIG.13. The homogeneousspherical earth model and its flat earth equivalent used
for the theoretical seismograms.

Thus for an homogeneous spherical model, we obtain the Cartesian model

The Legendre functions in the spherical wave solution are replaced by an asymptotic
expansion in terms of Bessel functions (Szego 1934). The leading term reduces to the
form of integral (11) except that an extra factor (elsin e)* is included (Muller 1971;
Gilbert & Helmberger 1972; Chapman 1973). Higher-order terms can be treated in
the same manner as the plane-wave expansion. The wave equation (1) must also be
transformed. If we transform the density according to the power law

the matrix A in (1) can be written as a finite series

Generalized ray theory

233

FIG.14. (a) Plane-wave, first-order seismograms for rn = - 5 , -3, - 1. Just the


component ux is shown. (b) As Fig. 14(a)except first-and third-order reflectionsare
included.

The matrix A' is identical to (2) and the other terms have been given in Chapman
(1973). Again we can convert equation (I) into (18) where B is replaced by a series as
in (FS). B(') is the function defined in (19) and
B(j)=sexp [-sQ]N-'A")Nexp [sQ] j = 1, ..., J .
(F6)
In general B(') can either be made zero by a suitable choice of m in (F4) (m = 3 for SH
waves or m = - 3 for acoustic waves) or can be included in B(') in modified y-functions
which are independent of m (Paper I). Higher-order terms (j= 2, ..., J ) contain
inverse powers of s and we have not considered their effect here.
To illustrate the effect of the density transformation (F4) we have evaluated
seismograms with various values of m without including B(') (Fig. 14). As this is the
fluid case, m = - 3 makes B(') zero and errors arise from B('). It is clear that for long
periods, the correct choice of m or the inclusion of B(') is as important as multiple
reflections (Fig. 4).

Das könnte Ihnen auch gefallen