You are on page 1of 13

Geophys. J . Int.

(1989) 99, 469-481

Efficient calculation of differential seismograms for lithospheric


receiver functions
G. E. Randall*
Departmen1 of Geological Sciences, SUNY at Binghamton, Binghamton, New York 13901, USA, and Earth Sciences Department, Lawrence
Livermore National Laboratory, Livermore, CA 94550, USA

Accepted 1989 April 17, Received 1989 April 17; in original form 1987 November 2

SUMMARY
A new technique for computing differential seismograms for crustal and upper

Key words: lithosphere, receiver functions, synthetic seismograms.


1 INTRODUCTION

Synthetic seismograms for the crustal and upper mantle


response to a teleseismic body wave (receiver functions) are
important tools for the interpretation teleseismic waveforms. Previously, receiver functions have been computed
either with ray theory for complex velocity models
(Langston 1977) or with the Haskell-Thomson formulation
(Haskell 1962) for laterally homogeneous layered velocity
models. Ray-theory synthetics can compute only specified
ray-paths, potentially missing arrivals and reverberations
whose importance may not have been anticipated. The
Haskell-Thomson technique can only compute a complete
seismogram of all arrivals; these synthetics can be as difficult
to interpret as the original observed seismogram, frustrating
attempts to explain the effects of changes in velocity models.
Furthermore, the Haskell-Thomson technique suffers from
numerical instability for frequencies and phase velocities
appropriate for teleseismic SV-waves propagating through
high-velocity layers typical of the lower crust and upper
mantle. Practical interpretation problems for teleseismic Pwaves (Zandt, Taylor & Ammon 1987) have shown the
difficulty and computational expense of specifying an
appropriate set of multiples in highly reverberant velocity
models, and in studies of teleseismic SV-waves (Zandt &
Randall 1985) the Haskell-Thomson technique was
frequently unusable because of the numerical instability.
* Current address: Seismological Laboratory, MacKay School of
Mines, University of Nevada, Reno, Reno, Nevada 89557, USA.

Although techniques exist for dealing with the numerical


instability of the Haskell-Thomson technique in surface
wave dispersion studies (Dunkin 1965) and reflectivity
studies (Kind 1978), none of these techniques are applicable
to the synthesis of receiver functions. A careful analysis of
Haskell's paper on receiver functions (Haskell 1962) shows
that the terms for the free surface displacement are not
composed of terms that are 2 x 2 minors of the original
matrix problem, as required by the methods that stabilize
the Haskell-Thomson technique.
Kennett (1983) has developed a technique for computing
the elastic wavefield in vertically stratified media. His
technique is numerically stable, and can synthesize a
pre-specified order of multiple internal reflections (including
a complete seismogram with all internal multiples) within a
laterally homogeneous layered structure. This paper
presents three formulations for synthetic seismograms of
layered receiver structures based on Kennett's technique.
The three formulations for synthetic seismograms are
used to develop an extremely efficient technique for
computing differential seismograms. Differential seismograms are used in linearized inversion, the process of
adjusting the parameters of a velocity model to create a
minimum mean square error fit between an observed
seismogram and a synthetic seismogram for the velocity
model. A first-order Taylor series approximation

is the basis for the linearized inversion technique (Menke

469

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

mantle response t o a teleseismic wave (receiver functions) is developed using the


matrix formalism of Kennett. The work was motivated by the difficulty of modelling
teleseismic SV-waves, and has also proven useful for modelling teleseismic P-waves.
This efficient method for calculating differential seismograms is based on three
separate methods for computing synthetic seismograms. Two of the synthetic
seismogram methods save intermediate results; then the remaining synthetic
seismogram algorithm uses the stored results in an efficient calculation of a new
synthetic seismogram for a perturbed velocity model. These developments have led
to a faster (for a 30-layer model, a 90 per cent reduction in computation time) and
more accurate linearized inversion scheme for the determination of velocity models
using teleseismic waves.

470

G . E. Randall

1984) that iteratively solves for perturbations to the


parameters of the velocity model. In the Taylor series,
Soh(t) is the observed seismogram, and S,,,[t, c u ( t ) ] is the
synthetic seismogram for a reference velocity model, ( ~ ( z ) .
The seismograms are treated as vectors of time samples, and
the Taylor expansion is taken about the unperturbed model
at each iteration of the inversion. The differential
seismogram,

SYNTHETIC RECEIVER FUNCTIONS

This section presents two direct formulations for the


computation of synthetic receiver functions with Kennett's
techniques, and an indirect formulation based on the result
of the two direct algorithms. The two direct formulations
compute the synthetic receiver functions by application of
Kennett's techniques, and require no results from prior
computations. The indirect technique requires that both of
the direct algorithms have been executed and saved
intermediate results, performing very modest computation
and indirectly doing the bulk of the computations by
referencing the previously computed results of the direct
techniques. For a forward modelling problem, the indirect
approach would be clearly impractical; however, the indirect
approach ,radically simplifies the computation of differential
seismograms. The indirect approach uses the intermediate
results from,the two direct approaches to compute a new
seismogram for a model with a single perturbed layer, and
need not recompute the intermediate results from the
unperturbed layers. The two direct approaches will be
discussed first, setting the stage for the indirect approach. In
the following discussions, the solutions will be in the
frequency-slowness domain, and can be inverse transformed
from the frequency domain to the time domain, at a fixed
slowness, to model a seismogram for a teleseismic arrival.

2.1 Direct bottom-up approach


The first technique presented is a straightforward bottom-up
approach. The description 'bottom-up' denotes the flow of
computations through the velocity structure. The idea is
simple, and the algorithm follows the energy through the
model. A transmission operator, TU, propagates the energy
up from the bottom of the structure to the base of the
surface layer. In parallel with the computation of T,, a
reflection operator R,, calculates the reflectivity of the
layered medium from the base of the surface layer to the
bottom of the model. The recursive computation of T, and
R, using Kennett's technique is outlined in the appendix.
These two operators are then combined with the free
surface reflection and diplacement operators to compute the
free surface displacement for a complete synthetic. This may
be represented by a simple matrix equation that can be
understood from right to left

d = W(I - plR~Np'R)-'p'T~Ni

(3)

where d is the vector of free surface displacements, and i is


the wave vector incident at the base of the model. The
model consists of N laterally homogeneous layers, with layer
1 bounded above by the free surface, and layer N is a
half-space. The matrix P1 is the diagonal matrix of phase
delays for propagation through the surface layer, as
described in the Appendix. The matrix W converts the
upward directed energy to free surface displacements. At
the free surface the reflection matrix is R. The construction
and interpretation of the matrices for the region bounded

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

represents the differential change in the seismogram for a


differential change of a single model parameter, or the
sensitivity of the seismogram to a single parameter in the
velocity model. In a typical inverse modelling study, the
computation of differential seismograms is a major burden,
equivalent to the computation of a perturbed synthetic for
each layer of the velocity model to be estimated, and one
synthetic for the original unperturbed velocity model. The
technique presented in this paper will compute differential
seismograms for a variation of a single parameter in each
layer of the model for slightly more than the cost of three
synthetics. For a velocity model of N layers, this results in a
total time of slightly over 3T, where T is the time for a
single synthetic, as opposed to (N + l ) T for a conventional
technique. The time for a single synthetic, T, is proportional
to the number of layers, N , so the efficient technique grows
linearly with N, but the conventional technique grows
quadratically with N. For a 30-layer model, a 90 per cent
reduction of computation time compared with a conventional approach represents a significant saving.
A major additional benefit of using t h e technique
presented here is an improvement of the accuracy of the
inverse modelling technique over using ray-theory synthetics. The WKBJ technique has been linearized (Shaw &
Orcutt 1985) for efficient waveform inversion of refraction
profiles, but for some receiver function modelling studies
ray methods are inadequate. Ray-theory synthetics in highly
reverberant velocity models are of necessity incomplete and
therefore inaccurate. Previously, the results of any inverse
modelling of teleseismic waveforms with ray-theory synthetics had to be appraised by computing a complete
Haskell-Thomson synthetic. In reverberant velocity models,
the Haskell-Thomson results frequently revealed the
problem of truncating the infinite reverberation series in a
finite ray-theory synthetic. Previous studies (Dr Steven R.
Taylor, personal communication 1986) had shown that a
brute force computation of differential seismograms with the
Haskell-Thomson technique was impractical when compared with ray theory. Although a previous study
(Fernandez 1965) had found a similar development for
computing differential seismograms based on the HaskellThomson technique, the report preceded the popular use of
inverse modelling techniques; the result was never published
in the general literature and the potential importance was
not widely recognized. The improved efficiency of the
technique described in this paper has made it faster and
more accurate to compute complete synthetic and
differential seismograms, and now ray theory is required
only when laterally inhomogeneous media are modelled.
An Appendix presents a summary of the relevant theory
of Kennett's technique, and examples of the application of

Kennett's method. This Appendix is included for those


readers who are unfamiliar with Kennett's work, but for a
complete treatment a careful reading of Kennett's
monograph (Kennett 1983) is recommended.

Differential seismograms for lithospheric receiver functions

-I

N-1

2.2 Direct top-down approach


The top-down approach alludes to the way the algorithm
proceeds from the top of the model to bottom, but is not
related to energy flow within the model. This technique is
less physically motivated, but still can be understood as a
sequence of reverberations.
The initial step consists of modelling the response of a
half-space by transforming from the incident wave vector to
the free surface displacements using the W operator, as
described in the Appendix. For convenience, the receiver
function for N layers will be denoted D',N and the net
reflectivity operator for a region of N layers bounded above
by the free surface will be denoted by RLN. For both D'.N
and RhN layer N is taken to be an infinite half-space. The
matrix DIyN is the operator that transforms the incident
vector of wave amplitudes i into the vector of free surface
displacements, d. The excitation, i, is applied at the top of
the half-space, just below the interface between layer N - 1
and the half-space layer N. The computations are initialized
for a half-space model with N = 1, by D'.'= W, and
RL1 = R, the free surface reflection operator.
The next step consists of calculating the response to a
single layer over a half-space. This is simply described by
D1,2 = Dl.l(I- plRb2p1Rbl)-1p1TL2

(4)

initializing the recursion for the receiver function, and

initializing the recursion for the reflectivity. The equation


for the receiver function represents the propagation of
energy through the interface at the bottom of the surface
layer, propagation of the direct arrival through the layer to
the free surface, and a sequence of reverberations. Each
reverberation consists of reflection from the free surface,
propagation down through the layer, reflection from the
base of the layer, and finally propagation back up to the free
surface to begin the sequence again. At the free surface,
each upward bound wave, either the initial direct wave or a
reverberati,on, is then transformed into free surface
displacements by the operator W. The matrix D',*
represents the net free surface displacement from upward
travelling energy incident at the bottom of the first layer. A
similar interpretation shows the reflectivity operator Rh2, to
be the net reflection from the first layer, accounting for all
reverberations within the layer.
Now both recursions have been initialized, and the
following two equations will recursively add layers to the
model. A schematic representation for the addition of a
layer to the structure is shown in Fig. 3. First the receiver
function is updated by

F i e 1. Schematic diagram for the reverberation operator in the


bottom-up method described by equation (3). This shows the direct
arrival and the first reverberation in the surface layer, using the net
transmissivity, T,, and net reflectivity, R,, for all layers below the
surface layer, computed by the technique discussed in the
Appendix. The free surface displacement operator is W,and R is
the reflection operator for the free surface. The propagation delay
through the surface layer is represented by PI.

where the reverberation now serves to compute the total


effect of the upward travelling waves at the top of layer
L - 1 caused by reverberation within layer L - 1. This
upward travelling energy is then transformed to free surface
displacements by D13L-',the receiver function for the L - 1
layer model. The net result of these operations is to treat
layer L as the new bottom half-space, and layer L - 1 is now

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

above by layer 1 and below by the half-space, RkN and


ThN, are discussed in the Appendix.
The usual reverberation operator interpretation is
attached to the (I-P'RkNP'R)-'
term, with a single
reverberation in the surface layer represented by PIRkNP'R
(see Appendix). Reading from right to left, this represents
reflection from the free surface, propagation delay down
through layer 1, reflection from the entire medium below
layer 1, and finally propagation delay back up through layer
1 to the free surface. A schematic diagram of this bottom up
technique is shown in Fig. 1, where a surface layer is
bounded above by a free surface and below by a layered
region with a net transmissivity and reflectivity calculated by
the methods described in the Appendix.
As stated in the Appendix, the reflection and transmission
matrices for the interfaces are independent of frequency for
elastic media, and only weakly dependent on frequency for
attenuating media. The frequency independence can result
in a substantial savings of computation because the interface
reflection and transmission matrices can be computed once
for a fixed slowness, and then used for all frequencies in a
synthesis. This saving can be applied to all three of the
synthetic receiver function algorithms.
This formulation was compared with the HaskellThomson formulation for accuracy and completeness. The
comparisons show that this bottom-up technique is
computationally efficient and numerically stable as predicted
by theory. Two comparisons of vertical component SV-wave
seismograms computed with Haskell's and Kennett's
techniques are shown in Figs 2(a) and (b) for different phase
velocities. All computations were for the P and S velocity
structure shown in Fig. 2(c).

471

472

G . E. Randall

Comparison of Kennett .vs. Haskell Synthetics

1.o-

0.5 0.0

-0.5
- 1 .o-

L
I

SP\

S\

I
I

20

I
I

60

40

1
1

80

Time (sec)
Figure 2(a). Comparison of vertical component of synthetic receiver functions computed with Haskell's and Kennett's methods using the
velocity model shown in Fig. 2(c). For comparison, both synthetics have been normalized to a maximum amplitude of unity. Both traces have
been low-pass filtered with a zero phase (non-causal) filter with the corner frequency at 1 Hz. Time is the time since the arrival of the incident
wave at the base of the model. These traces are computed for a phase velocity of 7.95 km s - ' , which means P-waves are evanescent in the
fourth layer (from 40 to 60 km) and SPr,,P is post-critically reflected from the Moho. The Sp phase shown is from the Moho.

a finite thickness layer connected to the bottom of the


previous results by the recursion formula. This process then
proceeds adding layers until all layers have been
incorporated in the result. In parallel with the recursion for
the receiver function, a similar recursion
RhL = T L - l s L p L - 1 1.L-1
D
Ru
1 L-1
L-1.L
(I - pL-IRn!,-1.LpL-'R1.#!-1
(7)
D
u
)- p T"
updates the reflectivity operator to include the reverberations within successive layers. The interface matrices
between layers L - 1 and L are TL-'3L, TL-lrL,R&-l,L,
and Rk-'9L as described in the Appendix. It is important to
note that RLL and DlSLdepend on layer L only through
these interface matrices.
The top-down technique may sound more complicated
than the direct bottom-up technique, but it is computationally equivalent. The description is more difficult because the
physical motivation for the algorithm is less obvious than the
motivation for the bottom-up approach. The computational
expense of D1,N and RhN is nearly identical to the
computational expense for RhN and TLN in the bottom-up
approach.

2.3 Indirect approach


The indirect approach requires that the intermediate results
from both direct approaches be saved before the indirect
algorithm can execute. This is clearly not a practical way to
compute a synthetic seismogram; however, it leads to a
practical technique for computing a set of differential
seismograms for a model with many layers.
The key to this technique is the description of the receiver
function using the results from the two direct techniques.
The top-down approach provides information about the
response above a given layer, and the bottom-up approach
provides information about the response below a given
layer. The remaining task is to combine these results from
the direct approaches with a reverberation sequence within
the layer to calculate a synthetic receiver function.
The receiver function formulation for the indirect approach
is based on a reverberation sequence for a chosen layer,
and the use of RhL and DISLfor the region above layer
L and the use of R k N and ThN for the region below
layer L. The equation for the synthetic receiver function is

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

-1.0

Dijferentiat seismograms for lithospheric receiver functions

473

Cornporkon of Kennett .vs. Haskell Synthetics

1 .o
0.5
Y

0.0

-0.5
-1.0
a

1 .o

0.5

-0
N

-0.5
-1.0
20

60

40
Time (sec)

80

Figure 2(b). The same comparison as in Fig. 2(a), with a phase velocity of 8.15 km
which means that P-waves can propagate through the
fourth layer. Both this figure and Fig. 2(a) show minor amounts of time domain wraparound caused by frequency domain aliasing. This can be
cured by taking a longer time window or using complex frequency to attenuate unwanted later arrivals.
s

and may be understood by expanding the reverberation


operator as discussed in the Appendix and examining the
individual terms. The expanded form of the equation for the
synthetic receiver function, with the direct arrival plus a
single reverberation is
D ~ , N

DI.LPLT$N

+D ~ . L ~ L R $ N ~ L R ~ L ~ L T G . N

(9)

where D1*N
is the synthetic receiver function for the total N
layer model. A schematic diagram of the direct arrival and a
single reverberation is shown in Fig. 4.
The first term from the expansion representing the direct
arrival, D'sLPLT$N,is simply understood from right to left
as the net upward transmission from the base of the N layer
model up to the base of layer L , followed by propagation
through layer L , and finally transformation into free surface
displacements by the receiver function for the region from
the free surface down to the top of layer L. The first
reverberation within layer L is represented by
D',LPLR$NP',RhLPLT~,N,
the second term from the
expansion. Reading from right to left again, term by term,
this is: the net upward transmission from the base of the N
layer model up to the base of layer L , followed by
propagation delay through layer L , reflection from the

'

region above layer L, propagation back down through layer


L , reflection from the region below layer L , propagation
back up through layer L , and finally transformation to free
surface displacements by the receiver function for the region
above layer L. The reverberation operator is the
representation for the sum of all the multiple reverberation
terms.

3 EFFICIENT COMPUTATION OF
DIFFERENTIAL SEISMOGRAMS
Differential seismograms can be easily computed by
carefully considering what parts of the indirect computation
are changed by the perturbation of parameters of the layer
of interest. Then, using the results previously computed by
the direct techniques for those parts of the computation left
unchanged by the perturbation of a single layer, the
seismogram for the perturbed model can be computed
without duplicating previous computations. The indirect
formulation is valid for any layer within the model, and the
computations for the two direct approaches can be modified
to save the intermediate results for every layer providing the
basis for the rapid computation of the indirect technique.

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

0.0

474

G. E. Randall
I

8-

6+
-

4 --

'

5-

&. 4c
Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

Depth (km)
Figure 2(c). P- and S-wave velocities versus depth for the crust and upper mantle model used for demonstration purposes.
D1,L-I

/"'"

RD

Schematic diagram of reverberation operators for the top


down case for D'.'- as described in equation (4). The figure shows
the addition of a layer at the bottom of a stack of layers in the
top-down progression. The direct arrival and a single reverberation
are schematically shown in layer L - 1, with R, representing the
reflectivity for the region above layer L - 1, and D',L-'
representing the receiver function for the region above layer L - 1.
Figure 3.

N-1

4. Schematic diagram of reverberation operators for the


indirect formulation of the synthetic receiver function calculation.
The direct arrival and a single reverberation are schematically
shown in layer L, with all operators as defined in Figs 1 and 3.
Figure

Differential seismograms for lithospheric receiver functions


When differential seismograms are computed, only a
single layer at a time is allowed to have perturbed
parameters, and some form of difference approximation is
used, such as

as,,,[t, +)I
adz)

- {sSy,[t,4
.
1
I

+ 6 4 z ) l - ssy"[4
+)I)

(10)

W Z )

where 6 a ( z ) represents the perturbed parameter in the


velocity model. If layer L is perturbed a careful analysis of
the receiver function shows that all the computations above
layer L (DlSLand R k L ) , and below layer L ( R g N and
TG.N), depend on layer L only through the interface
reflection and transmission matrices for the top and bottom
of layer L. This may be seen by examining equations (6) and
(7) for the region above layer L, and equations (A6) and
(A14) in the Appendix for the region below layer L.
Intermediate results saved in the bottom-up and top-down
computations are combined with the perturbed interface
reflection and transmission matrices to complete the

475

calculations D l S L ,RLL, R k N , and TG,N for the perturbed


model. These computations are less expensive than a normal
layer reverberation because they can use previously
computed partial results. The propagation phase delay
matrix for layer L also needs to be recomputed. Then the
indirect formulation can rapidly form the reverberation
using the modified layer L values. This provides a new
receiver function for the modified velocity model, and the
difference between this and the receiver function for the
unperturbed model represent the differential change in the
receiver function for a differential change in layer L. In Fig.
5 examples of perturbed and differential seismograms for a
perturbation of the velocity model of Fig. 2(c) (layer 4 was
perturbed by a 0.5 per cent P-wave velocity increase) are
shown.
For a single differential seismogram, the perturbed
receiver function is computed by using the saved values of
the transmission and reflection operators above and below
the perturbed layer from the direct techniques in a
Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

Comparison of Unperturbed .vs. Perturbed Synthetics

-kennett v 7.95
............. kennett vp 7.95

0.2

E
Q,

-s 0.0
.v)

-0.2
20

30

40

50

60

Time (sec)
Figure 5. Comparison of unperturbed and perturbed synthetics computed with Kennett's method, and comparison of the differential
seismograms computed with Haskell's method and Kennett's method. In the perturbed and differential seismograms, the variation of the
P-wave velocity in layer 4 of the velocity model (shown in Fig. 2c) is 0.5 per cent from 8.0 to 8.04 km s-'. All traces have been low-pass filtered
with a zero phase (non-causal) filter with the corner frequency at 1 Hz. The comparison of unperturbed and perturbed seismograms in 5(a) and
5(c) shows the effect of phase velocity. In 5(a) the phase velocity of 7.95 km s - ' means that P-waves are evanescent in layer 4, and minimal
energy tunnels up or down through layer 4 as a P-wave. The arrivals after about 45 s in 5(a) show amplitude changes, but minimal traveltime
change. In contrast, the phase velocity in 5(c) is 8.15 km-' s, and the later arrivals in 5(c) occur earlier because the slightly increased velocity in
layer 4 reduces the travel time for P-waves travelling through layer 4. The comparison of differential seismograms computed with Kennett's
and Haskell's methods are shown in 5(b) and 5(d) for phase velocities of 7.95 and 8.15 kms-', respectively. For the purposes of comparison,
all differential seismograms have been normalized. This demonstrates the equivalence of the two techniques.

Comparison of Kennett .vs. Hoskell Differentiol Seismograms

(b)

1.o
-L

0.5

g 0.0
V

0
-

-n
2
_ -0.5
-

-9
c
c

-1.0
1.0

YI-

c3
D

W
.-

0.5

-0.5
- 1 .o-

Time (sec)
Comparison of Unperturbed .vs. Perturbed Synthetics

(C)

0.2

c-'
I

20

30

....

40
Time (sec)
Figure 5. (Conrinued)

50

60

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

E 0.0

Diferential seismograms for tithospheric receiver functions

Comparison of Kennett

(d)

JS.

477

Haskell Differential Seismogroms

1.o

0.5

Q)

0.0

.-0.5
a
-

-= -1.0
-L

-n

1.0

Lc-

0.5

0.0

-0.5

-1.0
20

30

40
Time (sec)

50

60

Figure 5. (Continued)

computation for a perturbed layer reverberation. This


means the expense of computing a perturbed receiver
function for each of the N layers is nearly the expense for N
reverberation operators, which is nearly the same expense
for a single synthetic receiver function for an N-layer model.
This is the basis for the claim that the N differential
seismograms for an N-layer model can be computed in about
the time for three seismograms, when the cost for the two
direct computations are also considered. As discussed
earlier, the time required for a direct computation grows
linearly with the number of layers. Clearly as the number of
layers increases, this indirect formulation becomes more and
more attractive.
Additional computational reduction is achieved because
the perturbed reflection and transmission matrices for the
interfaces are independent of frequency for elastic media,
and only weakly dependent on frequency for attenuating
media. Thus, the perturbed reflection and transmission
matrices for the interfaces can be computed once for a fixed
slowness, and then used for all frequencies in a synthesis.
Computational efficiency has been achieved by increasing
the storage of intermediate results, and it is important to
note the storage required. All the computations discussed
above are for a single value of frequency and a single value
of slowness. Storage is required for the four matrices (DIsL,
RLL, RkN and T f y N ) for
, each of the N layers. Each matrix
is four complex numbers, typically requiring 8 bytes
per complex number, so N (layers) x 4 (matrices) x 4

(elements) x 8 (bytes) is required for a single frequency


and slowness. For a 32-layer model, the total storage for
intermediate results would be 4 kilobytes. Additional
storage for the perturbed interface reflection and transmission matrices is nearly twice the previous storage. There are
four complex 2 X 2 matrices for each interface, and these
must be computed for a variation in the layer parameters
above and below each interface, which results in additional
8 kilobytes for a 32-layer model. Some storage can be saved
by exploiting the symmetries for the transmission matrices
discussed in the Appendix. At worst, the total intermediate
storage would be only 12 kilobytes. The storage for the NF
complex frequency values of the synthetic and N differential
seismograms would easily be much more important than the
intermediate storage. For example, if N F is 1024 and N is
32, then the complex frequency domain storage is 33
(waveforms) X 1024 (frequencies) x 8 (bytes) or 264 kilobytes. For teleseismic receiver function modelling, only a
single slowness is typically required. If a spectral synthetic
with multiple slownesses were required, the order of the
numerical slowness integration technique would control the
number of stored frequency domain values at different
slownesses that had to be stored. A simple trapezoidal
integration or a first order Filon (Frazer & Gettrust 1984)
integration would allow the integral to be formed as a
running sum with no intermediate storage. In any case, the
storage for intermediate results would not grow.
The FORTRAN code implementing this theory runs on a

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

x
-

*-

478

G. E. Randall

4 DISCUSSION
This paper has described three techniques for computing
synthetic seismograms of the response of the lithosphere to
teleseismic waves. Each technique emphasizes reverberation
within different parts of the velocity model, and these
different viewpoints are used to exploit the intermediate
results of the first two techniques in a computationally
simple third technique. No single viewpoint for the
computation of the synthetic receiver functions is as
effective at minimizing the recomputation of results.
After the theory was developed, a computer program was
written to calculate differential seismograms, and the results
were verified by comparison with a brute force computation
using a Haskell-Thomson formulation. The dramatic speed
improvement motivated the incorporation of these algorithms into an existing inverse modelling code (Owens,
Zandt & Taylor 1984) that had been based on ray-theory
synthetics. The first results using the modified modelling
program have already been presented (Priestley , Zandt &
Randall 1988) for P-wave data.
The substantial reduction of computation time of
differential seismograms for the teleseismic modelling
presented here should serve as an incentive for other inverse
modelling studies such as refraction waveform modelling.
The multiple reformulation of the forward problem can
clearly reduce the computation time of differential
seismograms in complicated models involving many layers.
The analysis of intermediate storage discussed above should
apply to the refraction modelling case as well. A vectorized
algorithm (Phinney, Odom & Fryer 1987) would require
storage of intermediate results at every frequency, but the
interface matrices could be stored as frequency independent. For a synthetic with 1024 frequencies, the intermediate
storage would then be about 4 megabytes. A more thorough
analysis should be the topic of a paper specifically discussing
the linearization of the reflectivity problem.

CONCLUSIONS

The use of multiple synthetic seismogram formulations, and


storage of intermediate results, can be combined to create a
much faster computation of differential seismograms for
inverse modelling studies. The substantial speed increase
makes it realistic to appraise the non-uniqueness of
inversion solutions by comparing the results from a suite of
inversions with different starting models, and different
model parametrizations. The increased accuracy of a
complete synthetic, as opposed to a finite ray approximation, is a substantial additional benefit when laterally
homogeneous media are modelled.

ACKNOWLEDGMENTS
This research was supported at SUNY at Binghamton by
National Science Foundation grant NSF-EAR-8508125 and
in part by grant NSF-EAR4306562 for computational
facilities. Additional support was provided at Lawrence
Livermore National Laboratory through the Department of
Energy Contract W-7405-ENG-48.
I would like to acknowledge the careful review of the
work leading to this manuscript by Dr George Zandt and Dr
Steven R. Taylor. Dr Taylor pointed out the earlier work by
Fernandez. 1 would also like to acknowledge thoughtful
reviews of this manuscript by Charles Ammon, Dr Keith
Nakanishi, Dr Howard Patton, and Dr Norman Burkhard. I
would especially like to acknowledge the incisive comments
of an anonymous reviewer.

REFERENCES
Dunkin, J. W., 1965. Computation of modal solutions in layered,
elastic media at high frequencies, Bull. seism. SOC.A m . , 55,
335-358.
Fernandez, L. M., SJ., 1965. Spectrum of P Waves, p. 172, Saint
Louis University, St Louis, Missouri.
Frazer, L. N. & Gettrust, J. F., 1984. On a generalization of Filons
method and computation of oscillatory integrals of seismology,
Geophys. J. R. astr. SOC., 76, 461-481.
Haskell, N. A,, 1962. Crustal reflection of plane P and SV waves, 1.
geophys. Res., 67, 4751-4767.
Kennett, B. L. N., 1983. Seismic Wave Propagation in Stratified
Media, Cambridge University Press, Cambridge.
Kind, R., 1978. The reflectivity method for a buried source, J.
Geophys., 44,603-612.
Langston, C. A., 1977. The effect of planar dipping structure on
source and receiver responses for constant ray parameter, Bull.
seism. SOC.A m . , 67, 1029-1050.
Menke, W., 1984. Geophysical Data Analysis: Discrete Inversion
Theory, Academic Press, New York.
Owens, T . J., Zandt, G. & Taylor, S. R., 1984. Seismic evidence
for an ancient rift beneath the Cumberland Plateau, TN: a
detailed analysis of broadband teleseismic P-waveforms, J.
geophys. Res., 89, 7783-7795.
Phinney, R. A., Odom, R. I. & Fryer, G. J., 1987. Rapid
generation of synthetic seismograms in layered media by
vectorization of the algorithm, Bull. seism. SOC. A m . , 77,
2218-2226.
Priestley, K. F., Zandt, G. & Randall, G. E., 1988. Crustal
structure in Eastern Kazakh, U.S.S.R from teleseismic receiver
functions, Geophys. Res. Lett., 15, 613-616.
Shaw, P. R. & Orcutt, J. A., 1985. Waveform inversion of seismic
refraction data and applications to young Pacific crust,
Geophys. J. R . mtr. SOC.,82, 375-414.
Zandt, G. & Randall, G. E., 1985. Observation of shear-coupled P
Waves, Geophys. Res. Lett., 12, 565-568.

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

SUN 3/50, a modest 32-bit microprocessor based workstation with 4 megabytes of memory, and is easily contained
in memory without using virtual memory to dynamically
page the program. The SUN 3/50 had the optional floating
point coprocessor (MC68881). The synthetics (vertical and
radial), and differential seismograms (vertical and radial),
with respect to P-wave velocity in each of the layers were
computed for two models. In the first case the computations
were for a 27-layer velocity model using a sampling interval
of 0.025s (20Hz Nyquist), and 2048 time points. The
frequency domain conjugate symmetry for real time
functions is exploited to permit evaluation of the transforms
of the signals at 1025 frequency samples. This is a high
resolution example, and is excessively detailed for the
analysis of teleseismic data, demonstrating a loose upper
bound for computation time. The SUN 3/50 performed the
frequency domain computations in 1550 s and a SUN 4/280
performed the same computations in 170s. Both codes
required modest additional time for inverse Fourier
transforms and file i/o of the synthetic and differential
seismograms. A more typical computation, representing
parameters used for analysis of teleseismic SV waveforms,
with a velocity model of 18 layers using a Nyquist of 2 H z
and 512 time samples took 256 s on a SUN 3/50 and 28 s on
a SUN 4/280.

Differential seismograms for lithospheric receiver functions

479

a n d t , G., Taylor, S. R. & Ammon, C. J., 1987. Analysis of


teleseismic waveforms for structure beneath Medicine Lake
Volcano, Northern California, Seisrn. Res. Lett., 58, 34.

with the usual meaning for the superscripts. The matrix that
transforms P and SV amplitudes at the free surface into
physical displacements is

APPENDIX

W = ( wvp
~ R PwRS
wvs)

Introduction

where the second superscript denotes the incident wave type


and the first superscript denotes either vertical (V), or radial
(R) displacement at the free surface.
Kennett has chosen a normalization for wave vectors that
for a unit amplitude wave vector in a perfectly elastic
medium corresponds to unit energy flux in the depth
direction. This leads to symmetry properties for the
reflection and transmission matrices:

Reflection and transmission at an intedace


At an interface between two homogeneous layers, reflection
matrices are calculated that transform a wave vector of P
and SV incident amplitudes into reflected P and SV
amplitudes. In the notation for a reflection, a subscript D
denotes an incident wave vector from above the interface,
with energy directed downward and reflected energy
directed upward. Similarly, a subscript U denotes incident
wave vector from below an interface, with energy
propagating upward, and reflected energy directed downward. In Kennett's notation:

The superscripts in the matrix elements refer to the incident


and reflected wave type with the first letter being the
reflected wave and the second letter being the incident
wave. For example, an SP superscript represents an incident
P-wave, mode converted to an SV-wave by reflection.
A similar notation applies for transmission matrices with a
D subscript representing incident energy from above an
interface with incident and transmitted energy both directed
downward. Similarly, a U subscript denotes incident energy
from below an interface with both incident and transmitted
energy directed upward. Again the second superscript of the
matrix elements refers to the incident wave type, and now
the first superscript refers to the transmitted wave type. In
Kennett's notation:

rips.

R"")

Rgc = RgB + TGBPBREcPB(I - RGBPBR~cPB)-*~DE


(A6)

1.

Two additional matrices are needed for reflection at the


free surface and representation of physical free surface
displacements. The additional matrix for reflection of
upward directed energy at a free surface is also defined by
K

"=(RsP

Reflection and transmission in layered media


The next major area to develop involves the construction of
matrix reflection and transmission operators for a layered
region, not just a single interface. This involves combining
the results of the previous section with propagation through
layers for a net result that accounts for reverberation within
layers bounded by interfaces. The simplest cases are for a
single layer bounded above and below by infinite homogeneous half-spaces. From the results for these cases,
the general problems of multiple layers are solved by
recursion. Also, at this point the approximations for partial
reverberation are developed. A heuristic presentation will
be used to develop insight into the equations that Kennett
derives from a theoretical basis. Operators for reflection of
energy incident from above a layered region and
transmission of energy incident from below a layered region
are developed here, and used in the main body of the paper.
For a layer between two half-spaces denoted A, B, and C
from top to bottom and incident energy from above,
Kennett has shown that the reflection operator for the
region is

T,=[ T g p TLs
and T u = [ T ; T
~F
TS,PTF '
TCp T E

.GPP

where the superscript T denotes the matrix transpose


operation. These symmetry relations will also hold iR a more
general context. Computation and storage can both be saved
by exploiting these symmetry relations.
A point that may seem trivial now, but will simplify
reading of the following equations, is the notion that the
equations are understood from right to left. This is a
consequence of the matrix algebra and column vector
notation used by Kennett. The right to left notation is
already seen in the superscript notation for matrix elements,
with incident on the right and transmitted or reflected on the
left. Finally, the reflection and transmission matrices are
independent of frequency for elastic media, and only weakly
dependent on frequency for attenuating media. The
frequency independence can result in a substantial savings of
computation because the reflection and transmission
matrices can be computed once for a fixed slowness, and
then used for all frequencies in a synthesis. In the synthetic
seismograms for teleseismic waveforms, only one slowness is
needed but many frequencies must be computed, making
the savings significant.

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

This Appendix is intended to serve as a brief, heuristic


introduction to Kennett's notation and technique. The
interested reader is encouraged to consult Kennett's
monograph for detailed mathematical derivation of the
results .
The formulations presented below are all directed towards
solving the wave equation in horizontal plane layered elastic
or attenuating media. I have chosen to present only the
P-SV matrix formulation, and omit the simplification to an
analogous set of scalar equations for SH. The solutions will
be in the frequency-slowness domain, and can be inverse
transformed from the frequency domain to the time domain,
at a fixed slowness, to model a seismogram for a teleseismic
arrival.

480

G. E. Randall

where the matrix superscripts refer to the layers bounding


the interface for that matrix. The matrix P", is a diagonal
matrix of the vertical component of the delays for
propagation of P- and S-waves through layer B. This is
simply represented in the frequency domain using the
vertical slowness and layer thickness by the following:

corresponding to n = 1, are used to approximate the matrix


inverse, a simple interpretation appears involving a single
reverberation within layer B. Expanding,

where h is the layer thickness, w is the radian frequency, p


is the horizontal slowness, and the square root terms are P and S-wave vertical slownesses. respectively. The signs of
the square roots are chosen such that the imaginary part of
the square root is positive, which will result in a stable,
exponentially decaying representation for evanescent waves.
If the other possible convention for Fourier transform sign
was used, a correspondingly different convention for the
signs of the square roots would be required to ensure
stability.
The phase delay matrix is guaranteed stable by
construction, and this formulation causes Kennett's
technique to be stable even for high frequencies. This
formulation is possible because the overall energy transfer is
directed. For example, in transmission problems net energy
propagates in only one direction, either up or down, for
each problem solved. Intermediate reverberations temporarily reverse direction, but the net propagation is still
directed either up or down. Similarly, in reflection, the
direction of net energy propagation is always reversed. This
allows only phase delays to be considered, and they may be
constructed to be stable.
This is in contrast to the Haskell-Thomson approach that
uses a matrix propagator formalism that has both upward
and downward travelling waves, that must have different
phase delay sign conventions. In an evanescent wave
problem, one set of the phase terms will be stable, and the
other will of necessity be unstable. Theoretically, the
instabilities should cancel algebraically in the solution of
many problems, but numerically the problem is still unstable
because of finite precision arithmetic.
The expression for R'; is difficult to interpret directly but
fortunately a convenient expansion is available. The matrix
identity

The three terms in the sum on the right-hand side can be


analysed separately by reading each term from right to left.
The first term, RG", is just the reflection from the interface
separating layer A and B. The second term,
T;'UBPRREcPBFDB,represents transmission down through
the interface between A and B, propagation through B,
reflection from the interface between B and C , propagation
up through B, and finally transmission back up through the
interface between A and B.
The third term is more complicated and represents the
first internal reverberation within the layer B. The third
term, T " , H ~ R ~ g c ~ Brepresents
~~B~
transB~~c~
mission down through the AB interface, propagation
through B, reflection from the BC interface, propagation
through B, and now a reflection from the A B interface, and
another sequence of propagation through B reflection from
the BC interface, and propagation through B followed by a
final transmission back up through the AB interface.
These three terms combine to form an approximation to
the wavefield reflected from the region, here allowing only a
single internal reverberation in layer B. As higher order
terms are included in the partial sum approximation, more
internal reverberations are included in the approximate
wavefield. In the original expression using (1R$BPBREcPB)-' all the terms in the infinite series are
included, and therefore this term is really a reverberation
operator for the layer B, including all internal reverberations. A simple schematic diagram illustrating the
reverberation process is shown in Fig. Al(a).
A similar analysis for the operator representing

R;"

+ GBPBRgCPB(I+ RGBPBREcPB)GB (A12)

and collecting terms

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

(I - X)-l= I + x + X*(I- x)-l

R;'=

(A81

can be expanded repeatedly to produce terms of a matrix


geometric series, and a final remainder term. Providing the
series converges,

c X"
m

= (I - X)-l

"=O

an approximation to the result is then just the partial sum

I + x + x2+.. . + X" (I - x)-'


2=

where n is taken as large as required for an approximation


of a specified accuracy. The remainder term is then

Use of the partial sum here will correspond to partial


reverberation.
If the first two terms in the partial sum expansion,

Figure Al. Schematic diagrams for reflectivity and transmissivity.


In the top half the individual reverberations are shown for the
reflectivity as described in equation (A6) of the Appendix. In the
bottom half, the reverberations for the transmissivity as described in
equation (A14) of the Appendix are shown.

Diflerential seismograms for lithospheric receiver functions


transmission upward from C through B into A can be done.
A simple schematic for this reverberation sequence is shown
in Fig. Al(b). The complete form of the result is:

ec eB(lP ~ R ~ ~ P ~ R ; ~ ) - ~ P ~ T F . (A14)
=

The result for the corresponding expansion for a single


reverberation is

ec PuBPBTEc+ ~ u B P B R ~ c P B R < B P B T ~ c
=

(A15)

481

that only contains two terms that can be interpreted


separately. The first term, PUBPBTEc, is merely the direct
transmission from C through B to A ; the second is the term
corresponding to a single reverberation within layer B
before transmission into A. Here the reverberation
operator, (I - PBRgCPBRflB)-', is just the transpose of the
reverberation operator for the reflection case analysed
above. This relation can also be used to save significant
computations in each layer.

Downloaded from http://gji.oxfordjournals.org/ by guest on October 11, 2015

You might also like