Geophys. J. R. astr. Soc.

(1976) 47,151-177

Source shape estimation and deconvolution of

teleseismic bodywaves

Rob W. Clayton and Ralph A. Wiggins* Department of

Geophysics and Astronomy, University a/British Columbia, Vancouver, British
Columbia, Canada, V6T 1 W5

Received 1976 April 8

Summary. We consider the deconvolution of a suite of teleseismic recordings

of the same event in order to separate source and transmission path pheno-
mena. The assumption of source unif~rmity may restrict the range of azi-
muths and distances of the seismograms included in the suite. The source
shape is estimated by separately averaging the log amplitude spectra and the
phase spectra of the recordings. This method of source estimation uses the
redundant source information contained in secondary arrivals. The necessary
condition for this estimator to resolve the source wavelet is that the travel
times of the various secondary arrivals be evenly distributed with respect to
the initial arrivals. The subsequent deconvolution of the seismograms is
carried out by spectral division with two modifications. The first is the intro-
duction of a minimum allowable source spectral amplitude termed the
waterlevel. This parameter constrains the gain of the deconvolution ftlter in
regions where the seismogram has little or no information, and also trades-off
arrival time resolution with arrival amplitude resolution. The second modifi-
cation, designed to increase the time domain resolution of the deconvolution,
is the extension of the frequency range of the transmission path impulse
response spectrum beyond the optimal passband (the passband of the seismo-
grams). The justification for the extension lies in the fact that the impulse
response of the transmission path is itself a series of impulses which means
its spectrum is not band-limited. Thus, t.i.e impulse response is best repre-
sented by a continuous spectrum rather than one which is set to zero outside
the optimal passband. This continuity is achieved by a recursive application
of a unit-step prediction operator determined by Burg's maximum entropy
algorithm. The envelopes of the deconvolution are used to detect the pre-
sence of phase shifted arrivals.
Introduction
This paper examines the problem of the deconvolution of source functions from teleseismic
recordings. The first step in any deconvolution process is the estimation of the source time
function. The next section of this paper deals with this problem.
*Present address: Western Geophysical Company, PO Box 2469, Houston, Texas 77001, USA.
152 R. W. Clayton and R. A. Wiggins
A standard technique in exploration seismology is to decompose the source from the
seismogram autocorrelation function. This method assumes that the source is minimum
phase and that the impulse sequence behaves like white noise (Robinson 1967). Since
neither of these assumptions appeared to be generally valid for teleseismic recordings, the
technique was not used. Source estimation by homomorphic transformation has also been
suggested (Ulrych 1971), and this method is discussed in some detail. However, the low
quefrency assumption of this method was found to not be generally valid, and there is also
a problem of phase instabilities with the homomorphic transform itself.
Both of the above methods attempt to estimate the source from a single seismogram. We
decided that on the basis of a single recording of an event it would be difficult, if not impos-
sible, to devise a general method of estimating the source time function. For example, if an
earthquake source has more than one distinct motion, then it is difficult to decide whether
the effect as observed on a single seismogram belongs to the source or to the transmission
path. For this reason, we decided to restrict the problem to the case where a suite of seismo-
grams from the same event is available. The redundant source information contained in the
suite will allow us to estimate the source with greater confidence than if the estimation
were done on the basis of a single seismogram. Restrictions have to be placed on the distance
and azimuthal ranges of the suite of seismograms so that approximate source uniformity
can be assumed.
The actual deconvolution of the seismograms with a given source estimate is discussed in
Section 3. Techniques such as Wiener deconvolution, which convolve the seismogram with
an inverse of the estimated source, were not used because of the problems encountered when
the source is not minimum phase (Robinson 1967). In the frequency domain where decon-
volution is the division of the seismogram by the source, the problems with Wiener deconvo-
lution are seen to arise from the spectral locations where the source amplitude becomes
vanishingly small. The method of deconvolution by spectral division can be easily modified
to overcome this problem by constraining the minimum allowable source amplitude level
(Heimberger & Wiggins 1971 ; Dey Sarkar & Wiggins 197 6). A second modification which
allows increased time domain resolution of the impulse response is also outlined in that
section.
Finally, the deconvolution of a real example which employs the methods outlined is
presented in Section 4.

2 Source estimations
2.1 GENERAL ASSUMPTIONS

For teleseismic recordings there are no independent observations of the source time func-
tion, so it is necessary that it be estimated directly from the data. This, in general, will
present some problems because the source wavelet will usually have some characteristics
which overlap with those of the transmission path impulse response. Some assumptions will
be necessary to resolve such problems of non-uniqueness.
The general model that will be considered for the form of each seismogram recorded from
the same event is:
(1)

where xj(t) is the time series recorded at the j-th station, s(t) is the source wavelet, hi(t) is
the transmission path impulse response to the j-th station, nj(t) is noise, and * denotes
convolution.
 On teleseismic bodywaves 153
The assumption implicit in this model is that the source time function is uniform in the
sense that it is independent of both azimuth and takeoff angle. In general, this uniformity
assumption is not true, as the kinematic dislocation models of radiation indicate (Savage
1966), but its validity can be improved by restricting the azimuthal angles, epicentral dis-
tances, and time ranges of the suite of seismograms. Problems with this assumption may be
expected when surface reflected phases such as pP are mixed with main arrivals or if the
total radiated source shape is radically different for different takeoff angles.
We will assume that over a suite of seismograms the various arrivals are minor perturba-
tions of a uniform source. We seek to estimate an 'average' source which can be used in a
deconvolution process to remove the major source effects of all the arrivals. The procedure
usually produces visual enhancement of the recordings, which is one of the purposes of decon-
volution, even though the result may not be a true representation of the Earth's impulse
response. The criterion we used in judging the source estimators was the visual enhancement
that deconvolution produced. There is no adequate source-shape estimation theory at
present to allow the shape of a complicated estimated teleseismic source wavelet to be
related to the various observed earthquake parameters.
The source uniformity assumption is not sufficient to solve equation (I) for the source
wavelet. Additional assumptions about the features of either the impulse response or the
source wavelet are necessary. One common assumption that is made in exploration seismo-
logy is that the source is minimum phase. This allows a unique decomposition of the source
from its autocovariance function (Robinson 1967). If this assumption is coupled with the
assumption that the impulse response sequence is white, then the source may be obtained
by the decomposition of the estimated seismogram autocovariance (Robinson 1967). The
minimum phase assumption may be true for explosive sources, but it is not generally true for
earthquake sources. The example used in this paper appears to have a non-minimum phase
source. The assumption of the whiteness of the impulse sequence is also invalid for a single
teleseismic recording because, within a particular time window of interest, there are usually
only a few arrivals.
For the case when a particular arrival on the record is separated from its neighbours by a
time which is greater than the source length, then an obvious method of source estimation
is simply to pick this arrival as the source. Seismograms recorded near a distance of 30
degrees often have this feature (DeySarkar 1974; Dey Sarkar & Wiggins 1976). However, one
has to be confident that there is only one arrival in the particular time window, and that the
window contains all of the source. Also, if the source wavelet is not uniform, then this
estimate may not be the 'best-fitting' one.
A simple extension of this 'simplest arrival' method is the averaging of a particular arrival
over several traces after suitable time alignment. The essential assumption of this method is
that all other arrivals present within the time window defined by the source wavelet length
must average to zero. In other words, the other arrivals must have sufficiently different
phase velocities (moveouts) compared to the arrival which is to be enhanced. This method
uses the redundant source information in a particular arrival over the suite but not in the
secondary arrivals on the seismograms. The resulting deconvolution will necessarily be biased
toward the particular arrival on which the source estimate was based.
Another method of source estimation, which uses homomorphic transforms, has been
proposed (Ulrych 1971; Stoffa, Buhl & Bryan 1974). This method replaces the minimum
phase source and white impulse response assumptions by a low quefrency source assump-
tion. This assumption also appears to be unrealistic for complex earthquake sources. How-
ever, this method will be discussed in some detail because it forms the basis of the estimator
proposed in this paper.
154 R. W. Clayton and R. A. Wiggins
2.2 SOURCE ESTIMATION BY HOMOMORPHIC TRANSFORMS

The concept of a homomorphic transform from a convolutional space to an additive space

was first proposed by Oppenheim (1967), and was elaborated upon by Schafer (1969).
illrych (1971) has suggested that it is a suitable method for seismic source estimation.
The basic idea of the method is to transform the convolution operator of equation (1)
into an additive operator so that the system may be treated by linear filtering theory. The
transformation that accomplishes this and its inverse are given below (Oppenheim 1967).

x = H(x) = F- 1
{log [F(x)]} (2)

x = H- 1 (x) = F- 1 {exp[F(x)]} (3)

where H(·) denotes homomorphic transformation, F(·) denotes Fourier transform, and x
is the homomorphic transform of x.
The cepstrum x defined here is usually termed the complex cepstrum (denoting the usage
of the complex logarithm), to distinguish it from the cepstrum defined by Bogert et al.
(1963) which uses only the real part of the logarithm. Since the 'complex cepstrum' is the
only one used in this paper, it will be referred to as simply the cepstrum.
The phase information of x is preserved under the transform because the complex logar-
ithm is used. The homomorphic transform of equation (1 ), for the case when there is no
additive noise is:
x = H(x) = H(s *h) = H(s) + H(h) = f+ h (4)

Additive noise causes special problems under this transform because there is no simple
expression for the homomorphic transform of the addition operator. Additive noise pre-
dominately affects the phase part of the transform. This aspect will be discussed later.
To separate§ and h in equation (4) an assumption is needed about their cepstral relation-
ship. The usual assumption is that the source cepstrum shas a lower quefrency content than
the impulse response cepstrum iz (Schafer 1969; illrych 1971). Quefrency is the independent
variable of the cepstral domain (i.e. it is the 'frequency' content of the log spectra) and has
the dimension of time. If this assumption is true, then the source may be approximately
recovered by filtering out the higher order quefrency terms. This assumption is illustrated
by the example of Fig. 1. In this example the source is approximately recoverable from its
cepstral representation by retaining only the lower quefrency components, while the impulse
response is approximately recoverable by retaining the higher quefrency terms of its
cepstrum.
However, the low quefrency assumption is not considered to be sufficiently valid for
source estimation of teleseismic events. If the z-transform of the source function happens to
contain a zero near the unit circle, then the amplitude and phase spectra will change quite
rapidly in the vicinity of thid ~ero (Claerbout 1976). In the cepstral domain this translates
into higher order quefrency terms for the source. The amplitude spectra of a suite of 16
recordings of an event from Western Australia were averaged together to produce the ampli-
tude spectrum shown in Fig. 2. The presence of the hole (marked by an arrow) indicates
that there is a zero of the z-transform near the unit circle. Since this feature appears in all
of the amplitude spectra it is assumed to be a source effect and not part of the impulse
response. This hole makes contributions to the source cepstrum at quefrencies of 1.1 sand
higher. Since the impulse cepstrum is expected to have contributions in this range, the
source cepstrum cannot be effectively recovered by just filtering out higher quefrencies.
Several practical problems exist in the computation of homomorphic transforms. The
 On teleseismic bodywaves 155
SCJURCE IMPULSE RESPCJNSE

SCJURCE CEPSTRUM IMP. RESP. CEPSTRUM

SOURCE FILTF:R
t IMP. RESP. FILTER

_ ___jnL___ u
ESTIMATED S('J~f\CE ESTIMATED IMP. RESP.

Figure 1. An illustration of the low-quefrency ·source assumption. The cepstra of a simple source and
impulse are shown. The source cepstrum is low-passed with the filter shown while the impulse response
cepstrum is high-passed with the complement of the source filter. The result shows, that for this example,
the source is approximately a low quefrency process while the impulse response is a high quefrency
process.

AMPLITUDE SPECTRUM

0.0 1 .0 2.0 3.0.

FREQUENcY (HZJ
Figure 2. An example of a hole in the source amplitude spectrum. The amplitude spectrum shown is the
average of 16 seismograms of the West Australia event of 1970 March 24 recorded in the range 113° to
150°. The hole (marked by an arrow) is assumed to be part of the source rather than a transmission path
effect because it appears in all amplitude spectra of the suite at the same frequency position. The assump-
tion illustrated in Fig. 1 is invalid for this case because the hole contributes to the high quefrency terms to
the source capstrum.

main problem is the unwrapping of the phase curve. The complex logarithm is not unique
with respect to its imaginary component (the phase) so a factor of ±2mr, n =0, I, ... ,
must be added at each point to make the phase curve continuous (Schafer 1969). The usual
procedure for unwrapping the phase curve is to compare the difference between the phase at
156 R. W. Clayton and R. A. Wiggins
a particular lag with the phase at the preceding lag. If this difference is greater than 1f, then a
±2n factor is added to the phase to make the difference less than n (Schafer 1969). This
procedure is started at zero frequency where the true phase is known to be zero. If the time
series has a negative DC component, then it is inverted before the transform is taken
(Schafer 1969). A linear phase component, which corresponds to a shift in time of the signal,
is removed from the unwrapped phase curve to prevent the linear component from dominat-
ing the cepstrum.
Additive noise can make the unwrapping procedure unstable for two reasons. The first
reason is that the phase deviation due to additive noise can become large when the signal
amplitude is low. To illustrate this, suppose that a particular sample of the signal spectrum
has Gaussian noise added to its real and imaginary parts. Fig. 3 illustrates two cases. When
the signal amplitude is high, the phase deviations remain at a reasonable level. However,
when the signal amplitude drops, the phase deviation becomes quite large. In the case where
the signal amplitude drops to zero, the amplitude deviations assume a Rayleigh distribution
and the phase has a uniform distribution over [-n, n] -Bracewell (1965, p. 335). Thus, for
low amplitude regions of the signal spectrum, the phase can have a large variance.

lm lm

~phase
\ \viations

eS

Re Re
Low Signal Amplitude High Signal Amp I itude
Figure 3. The effect of additive noise on amplitude and phase. The circle surrounding the signal point S
approximates one standard deviation of additive noise. The corresponding amplitude and phase deviations
are shown. As the amplitude of S becomes small, the amplitude deviations remain constant but the phase
deviations become large.

The second problem also occurs at low amplitude regions of the spectrum. A hole in the
amplitude spectrum indicates the presence of a zero in the z-transform near the unit circle.
Suppose that the true signal has a zero just inside the unit circle and that additive noise
forces it just outside the unit circle. The phase of the true signal would increase by n near
the zero, but when the noise is added the phase decreases by n at this point. Fig. 4 shows
this effect on the unwrapped phase as well as on the unwrapped phase with the linear trend
removed. It is obvious that a change in location of zeros near the unit circle can drastically
alter the shape of the unwrapped phase curve with the linear component removed.
Several methods were tried in an attempt to remove the phase instability at low ampli-
tudes. Although none of them worked sufficiently well to be deemed useful, they are listed
here because the problem is usually ignored in the literature.
I AMPLITUDE SPECTRUM
On teleseismic bodywaves
L~CATI~N ~F ZER~S
lm
157

®4
®3
' 2.4
' , RAD.
''
' Re

UNWRAPPED PHASE L.P.C. REM~VED

4rr 1&3
=-=-= 1 &4
----- 2&3
1T
/"
------ \
-+-+2&4 I

2rr

0
0.0 1.0 2.0 3.0 1.0 2.0 3.0
FREQ. CRAD./SECJ FREQ. CRRD ./SECJ
Figure 4. The effect of changing a zero location on the phase curve. The presence of a hole in the ampli-
tude spectrum of a function indicates there is a zero near the unit circle in the z-transform of the
function. If additive noise causes one of these zeros to move across the unit circle line, then the phase
deviations become large. This figure illustrates the four diffe~ent phase curves for a pair of zeros moved
in this fashion. The amplitude spectrum of the four phase curves is the same. The phase deviations are
more pronounced when the linear phase component (LPC) is removed as is done with homomorphic
transforms.

(1) The phase was estimated by interpolation from its neighbouring points when the
amplitude dropped below a certain level. Two problems make this method impractical.
First, it is difficult to define a cutoff level which rejects all poorly-estimated phase term~
while retaining a sufficient number of well estimated ones to make interpolation worthwhile.
Secondly, the method tends to remove all n jumps, even the ones that should be present.
(2) Stoffa et al. (1974) suggested computing the derivative of the phase directly from the
real and imaginary parts of the F'lurier transform and then integrating to obtain the
unwrapped phase curve. If the Fourier transform is X ( w) + i Y ( w ), then the phase derivative
is: (Stoffa et al. 1974).

<I>' (w) = [Y(w) X' (w)- X(w) Y(w')] /A (w) 2 (5)

where A (w) is the amplitude spectrum. The phase derivative also becomes unstable at low
amplitudes because of the A 2 ( w) term in the denominator. Also, integrating the phase deri-
vative in regions of rapid phase changes can be a problem, because of poor sampling of the
phase derivative. Interpolating the phase derivative over low amplitude frequencies tends to
remove alln jumps, as with the first method.
Essentially, the problems with the phase unwrapping techniques reduce to the following.
While it is possible to unwrap the phase in regions of high spectral amplitude, there appears
to be no way of connecting the unwrapped sections of the phase curve across regions of low
spectral amplitude.
158 R. W. Clayton and R. A. Wiggins
We conclude that the filtered cepstrum is not a good source estimator when it is used on a
single recording. The phase estimate is unstable in the presence of noise and there is doubt of
the general validity of the low quefrency assumption.
More successful results are obtained by averaging the log amplitude and phase spectra of a
number of records from the same event. This tends to make the phase more stable and it
frees the method from the low quefrency assumption. This estimator will be discussed in the
next section.

2.3 SOURCE ESTIMATION BY AVERAGED LOG AMPLITUDE AND PHASE SPECTRA

A method of'averaged log amplitude and phase spectra is proposed as a technique for freeing
the single trace homomorphic transform from the low quefrency assumption and for improv-
ing the phase stability. Since the expectation operator is commutative with the Fourier
transform, averaging the log amplitude and phase spectra is equivalent to averaging the
cepstra. When the phase is averaged, the raw phase and not the unwrapped phase is used, so
that the instabilities of the unwrapping procedure can usually be avoided. The object of the
averaging is to enhance the source part of the cepstrum while diminishing the impulse
response part. The fundamental assumption made here is that although an individual seismo-
gram has non-random arrival times, the arrival times in a suite of seismograms considered
together are sufficiently random for cepstral averaging to reduce the impulse response effect.
If the averaging succeeds in eliminating most of the impulse cepstrum, then the low que-
frency assumption need not be applied to recover the source.
The impulse response of an ideal earth (a non-attenuating, non-dispersive, and non-phase-
shifting earth) is a sum of Dirac impulses.
N
ht = L
n~l
Cn 8 (t - Tn) (6)

where Cn is the amplitude of the n-th arrival and Tn is its arrival time.
Attenuation and dispersion preferentially modify the frequency components of the
source wavelet as it travels through the Earth. Any source estimation method which
attempts to find an 'average' source wavelet over a suite of seismograms will out of necessity
also include an 'average' attenuation and dispersion factor. We will consider attenuation and
dispersion phenomena, apart from the 'average' effect which will be included with the source
wavelet, to be minor perturbations on the form of equation (6). If phase-shifting were
included in the impulse response model, then a phase term dependent on the summation
index would have to be included in the sinusoidal terms of the following three equations.
The omission of this effect does not affect the arguments based on these equations. The
principal effect of variations in the phase of various arrivals is that the estimated source may
have a uniform phase delay. The subsequent use of the envelope of the deconvolved traces
nullifies the effects of uniform phase shifts.
The Fourier transform of the impulse response model is
N
H(w) = L en exp(-iwTn) (7)
n~l

The log amplitude (logA(w)) and phase (<I:>(w)) of H(w) are

= lhlog Lt
logA(w)
1 C~ + 2 ntl m~n CnCm cosw(Tn- Tm)} (8)
I <l>(w) = tan- 1
ntl Cn sin WTnl
On teleseismic bodywaves 159

(9)
[I
N

n=l
Cn coswTn

The log amplitude will have its main quefrency contributions at all combinations of the
differences of the various travel times. The average of the log amplitude spectra over a suite
of impulse responses will tend to a constant, as the number of members in the suite is
increased, if the combinations of travel time differences are sufficiently uniformly distri-
buted among the impulse responses. In the time domain, this means that the various arrivals
must have differing phase velocities (moveouts). The constant term
N
I c~
n =l

of the log amplitude is the average of the energy in each impulse response. This will intro-
duce a multiplicative constant into the source estimate which is not important for decon-
volution. The phase spectra will also have contributions which correspond to the various
travel times so that an average of the phase over a suite of impulse responses will tend to
zero under the same conditions as the amplitude spectra. In general the impulse response
components of the seismogram cepstra will not average to zero, but if they are reduced
sufficiently, the source wavelet will be enhanced when the cepstrum is transformed to the
time domain. It is usually necessary to apply a time window to the source wavelet to remove
any remaining effects of the impulse response.
To form the source estimate, the log amplitude and phase spectra of a number of records
are averaged together. We found that the best results were obtained, as would be expected,
when noisy records or rec~rds with obviously different sources were excluded. A sufficient
number of 'good' recordings must be retained to make the averaging process effective. Some
care is necessary when the phase is averaged because of the arbitrary branch cut in the arc-
tangent function at ±1r. This break would only be suitable if the expected value of the source
phase is zero. However, if the expected value is not zero, then the branch should be placed 1T
radians above and below this expected value. An initial estimate of the phase is necessary to
establish the correct interval over which to average. Each phase value is then placed in this
interval by adding the appropriate ±27T, if its difference from the initial estimate is greater
than 7T.
The initial phase estimate is taken as the phase value whose corresponding spectrum is
closest to the average amplitude spectrum at each frequency sampling point. This is the
phase that is most likely to be independent of the phase spectrum of the impulse response.
It is necessary that scale factor differences among the amplitude spectra be removed
before the averaging in order that the initial phase estimate is not biased by energy differ-
ences in the seismograms. Rather than normalizing the trace amplitudes by the peak trace
height or the peak spectral amplitude (both of which could be strongly affected by the
impulse response) a multiplicative constant for each trace was determined by least squares.
The first trace was picked as a reference and the following least squares problem was solved
for the {Cj}.
2
e == L {A 1(w)- qAj(w)JZ (10)
w

where A i (w) is the amplitude spectrum of the j- th recording.

160 R. W. Clayton and R. A. Wiggins
The summation is taken over the frequency passband of the recordings. The solution of
0 which minimizes the error is
q =L Al(w)Aj(w)jL A](w) (11)
w w

The amplitude spectra of all other traces are then normalized by the 0 factors.

Aj(w) = Ai(w) · Ci j = 2, ... ,N (12)

After averaging the log amplitude spectra and phase spectra, the low quefrency source
assumption may be applied, if it is desired. To do this, the final inverse Fourier transform of
the homomorphic transform is taken and the resulting cepstrum is zeroed above some que-
frency level. However, as mentioned earlier, the low quefrency assumption is a dubious one
for complex earthquake sources. The assumption, if it used at all, must therefore be applied
with caution. Furthermore, since the impulse response is diminished by the averaging pro-
cess, there will generally be no clear indication of the correct upper quefrency limit.
If there is a secondary arrival on the seismograms that does not move out significantly
with respect to the primary signal, then this method has trouble separating the source from
the combination of the two arrivals. If a source is estimated and it appears to have two
separate parts, then a decision must be made whether the source estimate actually contains
two separate arrivals or whether both parts belong to the source. The easiest method of
deciding this question is to check f0r the presence 0f holes (cf Fig. 2) in the amplitude spec-
trum of each recording. The frequency of a hole corresponds to the time separation of the
two apparent components. If this feature is present on all records and its frequency position
does not change, then it may be concluded that, within the resolving power of the suite of
seismograms, both parts belong to the source. Examining the amplitude spectra in this man-
ner is more reliable than looking directly at the records themselves for apparent moveout
because; first, the feature is usually more obvious in the amplitude spectra, and second,
the amplitude spectra uses the redundant information of all the arrivals on the record.
Several similar source estimators can be formulated along the same lines as the one given
above. For example, a simple average of the amplitude spectra could be substituted for the
average of the log amplitude spectra. The median of the amplitude spectra could also be
used. The arguments given earlier that the impulse response contributions will cancel out if
the various arrivals on the seismograms have different phase velocities also apply to these
estimators. If the source wavelet shape is quite uniform over the whole suite of seismograms
then the performances of the estimator outlined in this paper, and the two mentioned above
are similar. However, we have found that when a non-uniform source is present, deconvolu-
tion with the source estimated by average log spectra and phase spectra gives a superior
enhancement of the seismograms.

3 Deconvolution of the seismograms

A method will now be outlined for the deconvolution of the individual seismograms with a
given source estimate. The desirable properties of a deconvolution method are: (1) it resolves
the arrival times sharply, and estimates the arrival amplitudes accurately, (2) it should be
stable with respect to small source estimation errors or source non-stationarity, and (3) it
should be robust with respect to the rejection of random noise.
The basic method that will be used is that of divisional deconvolution in the frequency
domain, because it does not require assumptions about the form of the source wavelet (i.e.
I On teleseismic bodywaves
that it be minimum phase). Two modifications will be made to improve the stability and
resolution of this method.
161

The first modification will be the introduction of a minimum allowable source amplitude
level (Heimberger & Wiggins 1971; Dey Sarkar 1974; Dey Sarkar & Wiggins 1976), termed
the waterlevel, to reduce spurious noise components and the effect of small errors in source
estimation. The second modification is the bandwidth extension of the impulse response
beyond its optimum passband, to improve the time domain resolution of the deconvolution.
The last part of this section deals with the envelope of deconvolution, which provides a
useful tool for the detection of phase-shifted arrivals.

3.1 THE WATERLEVEL PARAMETER

The frequency domain form of the seismogram model- equation (1)- is

(13)

where the capitals denote the Fourier transform pairs of the quantities in equation (1 ). To
obtain the estimate of the impulse response Hi, the spectrum of the seismogram is divided by
the estimated source S.

(14)

where * denotes conjugation.

As the estimated source amplitude becomes small (IS i --'; 0), the factor multiplying~ is
0(1), assuming that S does not deviate too far from S. However, the factor multiplying the
noise component~ is 0 (1/1 S /). Therefore, it is essential to establish a minimum amplitude
level for the source to prevent the noise term from becoming too large. This level can be
thought of as a limit on the gain of the filter 1/S in the spectral regions where the seismo-
gram has little or no information. The minimum source amplitude is termed the waterlevel
and is conveniently expressed as a fraction of the maximum source amplitude. With this
modification the estimator of H becomes

-* -*
SS ~+~S
(15)

where k is the waterlevel parameter (0< k< I) and /Simax is the maximum source amplitude.
The deconvolution will be stable with respect to small errors in the source estimation if
the factor

' ss-* (16)

does not become unstable as a function of frequency. To show the effect of the waterlevel
parameter on this type of error let the real source be

S = S + cR (17)
6
162 R. W. Clayton and R. A. Wiggins
where R is an arbitrary function of unit maximum amplitude and c is a scale factor which
indicates the degree of error in estimation. The factor (16) now becomes

(s + cR) S* ISI 2 + cRS*

(18)
max {IS 12 , k 2 1S l~ax} max {IS 12 , k 2 1S l~ax}

As with the noise, the factor multiplying cR is 0 (1/1 S 1). Therefore, the introduction of the
waterlevel also prevents this factor from becoming too large, if cis sufficiently small.
An optimum choice for k would be one which makes k IS Imax greater than both the
noise level and the deviations of the estimated source from the real source. Since these fac-
tors cannot be determined in practice, an alternate procedure is suggested. Rather than
attempt to pick an optimum k, we suggest that the deconvolution be performed for a range
of k E [0, 1] . The stability of the deconvolution can be checked by comparing the impulse
response for the various waterlevels. This is the procedure that is used for the examples in
this paper.
The waterlevel parameter has another interesting interpretation. As k approaches zero,
the estimator becomes an unrestricted de::;onvolution of Xj by S. As k approaches unity, the
estimator is just a scale factor times the crosscorrelation of~ and S. The unrestricted decon-
volution attempts to remove all of the source effects from the seismogram, thus making it
the best estimator of the true impulse response. This form of the estimator will be the best
for resolving the travel times. The crosscorrelation is the least squares estimate of the arrival
amplitudes (Heimberger & Wiggins 1971). If a particular phasearrives at time T, then the
squared error due to amplitude differences between the arrival and the estimated source is

(19)

The amplitude factor A which minimizes this error is

A = L s(t) x(t- T)/L s(t) 2

(20)
t t

which is just a scale factor times the crosscorrelation of x(t) and s(t), evaluated at timeT.
If s(t) is normalized to unit maximum amplitude, then A is the arrival amplitude. To obtain
A from the divisional deconvolution with k = 1, the peaks are multiplied by the maximum
source spectral amplitude squared and divided by the energy of the source. The waterlevel
can therefore, also be interpreted as a parameter which trades off amplitude resolution with
arrival time resolution. This is added incentive to perform the deconvolution for a range of
waterlevels.

3.2 EXTENSION O.F THE IMPULSE RESPONSE SPECTRUM

The second modification that was introduced into the divisional deconvolution method was
an impulse response spectral extension scheme designed to increase the time domain resolu-
tion of the estimated impulse response. This scheme consists of assuming a model for the
impulse spectrum and predicting the unknown spectral components on the basis of this
model. In the section on source estimation it was shown that the assumption of source
stationarity led to an impulse response which was a sum of Dirac impulses for a non-disper-
I sive, non-phase shifting earth
N
On teleseismic bodywaves
equation (6). The Fourier transform of this relation is
163

H(w) = I Cn exp {-iwTn} (21)

n=l

which is just a sum of N complex sinusoids.

When divisional deconvolution is performed, it is necessary to ftlter the impulse response
to a passband defined by the band of significant energy of the particular seismogram. If this
were not done, then spurious noise information could be included in the impulse estimate as
is shown in the example of Fig. 5. The large spikes atl.l Hz and 1.4 Hz are noise compon-
ents which are enhanced by divisional deconvolution. The ringing nature of the unftltered
impulse response is due to these spikes. When the impulse response is ftltered with the indi-
V· ~ated passband, the result is much more stable. This approach, however, makes a zero exten-
sion assumption about the nature of the impulse response outside the passband and, in view

SDURCE RMPLlTUDE SPECTRUM

PL
IMPULSE RMPLITUDE SPECTRUM
----------------------

0.0 1.0 2.0 3.0

FREQUENCY lHZJ

UNFILTERED IMPULSE ReSPONSE

FILTERED IMPULSE RESPONSE

0. I 0. 20. 30. 40.

TlME l5ECJ
Figure 5. The effect of choice of passband on the impulse response. This example is the AYSD recording
of the Kern County event. When the deconvolution is not restricted to a passband of significant source
energy, the estimated impulse response is dominated by the enhanced noise spikes. The result is much
more stable when the deconvolved impulse response is filtered by the indicated passband.
164 R. W. Clayton and R. A. Wiggins
of the model of the impulse spectrum- equation (21)- this does not seem to be a particu-
larly good assumption. A better assumption would be one which fills in the impulse
spectrum in accordance with the continuous nature of this :nodel.
A procedure for extending short .realizations of a continuous process by the maximum
entropy method (Burg 1967), has been given by Smylie, Clarke & Ulrych (1973) and Ulrych
& Bishop (197 5). A prediction filter is found by fitting the following autoregressive process
to the known data

p
L akxi-k + zi
k=l
(22)

where. {ad are the coefficients of the prediction operator, p is the order of the autoregres-
sive process (the length of the prediction operator), and zi is a white noise process termed
the innovation. ·
An alternate way of writing the autoregressive model is in terms of the prediction error
operator:

p
L ·lkXj-k = Zj (23)
k=O

where "Yk =- ak, k = 1, ... ,p, and ro= 1.

For Burg's maximum eQtropy algorithm the process is fitted by minimizing the variance of
the innovation when the prediction error operator is applied in both the forward and reverse ·
directions (Smylie et al. 1973). The method will always produce a minimum phase predic-
tion error operator (Burg 1967; Claerbout 1976).
For the problem of extending the frequency range of the impulse response, the {xi} of
equations (22) and (23) is the Fourier spectrum of the impulse response. The minimum
phase property of the prediction error operator is ideal for extending the impulse spectrum
because the energy of the predicted part of the impulse spectrum will not increase away
from the known passband. Since the impulse spectrum is complex, the complex form of
Burg's algorithm will be required to compute the prediction operator. The derivation of the
complex Burg algorithm follows along the same lines as the real version, except that it must
be realized that the prediction operator is conjugated for reverse prediction (Smylie et al.
1973).
It could be argued that since the impulse spectrum model is a sum of sinusoids, it would
be better to fit a sinusoidal model rather than an autoregressive model to the known data.
The problem with this approach is that the sinusoidal model can only be assumed to be
approximately valid, due to attenuation and phase shifts of the arrivals. The autoregressive
model makes the more general assumption that the given impulse passband is part of a con-.
tinuous process which covers the entire spectrum. If the spectrum is truly a sum of sinusoids,
the autoregressive model does an excellent job, as long as there is some additive noise present
(Ulrych & Bishop 1975). ·
Once the prediction operator has been determined by the maximum entropy method, it
is applied in a. unit step prediction fashion, until the entire spectrum is filled in. An example
of spectral extension is shown in Fig. 6. The trace shown in Fig. 6(a) was generated by band-
passing an impulse sequence consisting of three impulses. The real part of the Fourier trans-
form of this trace is shown in Fig. 6(b ). The real part and its associated complex part of the
Fourier transform inside the passband were used to construct a prediction operator which
was then used to extend the spectrum. The real part of the extended spectrum is shown in
 On teleseismic bodywaves 165
A

_lJ_r-··- -
Figure 6. Spectral extension applied to a simple impulse response. (A) A synthetic trace consisting of
three impulses which was bandpassed with an ideal filter to reduce their resolution. (B) The real part of
the Fourier transform of (A). A complex prediction operator which was fitted to the part of the trans-
form within the passband, was used to extend the spectrum by unit-step prediction. (C) The real part of
the extended spectrum. (D) The result of transforming the extended spectrum back into time domain. A
comparison of (A) and (D) indicates the increase in resolution that is achieved.

Fig. 6(c). When the extended spectrum is transformed back into the time domain (Fig. 6(d)),
the increase in resolution of the impulse sequence is obvious.
The amplitude of the predicted oscillations of the spectrum always decreases away from
the optimal impulse passband because: first, the prediction error operator isminimum phase,
and second, the innovation of the autoregressive model is not included in the prediction.
The one parameter that has to be chosen for this scheme is the length of the prediction
operator. The same problem occurs in the estimation of power spectra by the maximum
entropy method (Ulrych & Bishop 1975). One criterion. that has been suggested for opti-
mally choosing this parameter is that of the final prediction error (Akaike 1969; Fryer,
Odegard & Sutton 1975; Ulrych & Bishop 1975). This criterion constructs a trade-off curve
between the minimum of the innovation variance and the variance of the estimate of the
prediction coefficients (Akaike 1969). The minimum of this curve is taken as the optimal
autoregressive order. However, observational experience indicates that for sinusoidal type
processes, significant resolution can be obtained by increasing the autoregressive order
beyond that indicated by the final prediction error criterion. The penalty that one pays for
doing so is the possible introduction of spurious noise components into the estimate. Instead
of attempting to find an optimum autoregressive order, we decided to use the same pro-
166 R. W. Clayton and R. A. Wiggins
cedure as with the waterleve1 parameter. That is, the deconvolution was performed for a
number of prediction lengths and the stability of the deconvolution is checked by comparing
the various results.
The impulse response passband for the calculation of the prediction coefficients was
determined by first performing a preliminary deconvolution in a passband determined from
the main spectral energy of the source estimate. If any large noise spikes were found to be
present, the passband was reduced to exclude them. Fig. 5 shows an example of these noise
spikes and the passband chosen to exclude the large spikes.

3.3 ENVELOPES OF DECONVOLUTION

Several authors have demonstrated that phase shifting of arrivals can introduce systematic
errors into travel time and amplitude observations by changing the relative peak and trough
positions of the arrival wavelets (Choy & Richards 1975; Hill 1974; Heimberger & Wiggins
1971). A simple procedure to overcome this problem is to compute the envelope of the
deconvolution (Heimberger & Wiggins 1971; Farnbach 1975).
The envelope is defined as the modulus of the analytic signal (Bracewell1965, pp. 267-
272) and can easily be shown to be independent of phase shifts (Clayton, McClary & Wiggins
1976). A useful way of displaying the envelope to show the phase properties of the signal is
to superimpose it on the deconvolution. Fig. 7 shows the typical relationships between an
impulse and its envelope for various phase shifts. A 90 degree phase shift, for example, is
characterized by a trough followed by a peak of equal amplitude.

PHRSE SHIFTS:
0 45 90 135 180

Figure 7. The relationship between a phase-shifted impulse and its envelope. Several band-passed, phase-
shifted impulses are shown with their superimposed envelopes, for various choices of phase shift angle.
The envelope may be used as a reference for gauging the phase shift angle of a particular impulse because
it is independent of the phase shift angle.

4 An example of deconvolution
The source estimation and deconvolution techniques outlined in the previous two sections
have been applied to some of the LRSN recordings of the Kern County, California earth-
quake of 1962 September 16. This event was used by Heimberger & Wiggins 1971 and
Wiggins & Heimberger 1973. The suite of seismograms is shown in Fig. 8.
This example has an interesting source wavelet which invalidates the assumption of low
quefrency sources. It is apparent on most records, particularly those beyond 29° epicentral
distance, that the source has a precursor feature about 5 s before the main event. If the low
quefrency assumption were applied to this source, then the precursor would be taken as part
of the impulse response, and not part of the source. It is also apparent by examining HTMN
and HNME that some relative phase shifting takes place between these two parts of the
source, or that the Earth's impulse response contains a phase shift.
 On teleseismic bodywaves 167
KERN 16/9/62
LCNM
10. 12

PMWY
11.34

12.22

, - - . - - - 15.64

15.67
~------------------

16.51

18.81

19.81

'v---
20.41

21.16

o. 10. 20. 31. 41.

Figure Sa.

Figure 8. The seismogram suite for the deconvolution example. This suite of seismograms is a set of LRSN
recordings of the Kern County event of 1962 September 16. The numbers to the right refer to the epicen-
tral distances of the stations, and the asterisks indicate the seismograms that where used in the source
estimate.
168 R. W. Clayton and R. A. Wiggins
KERN 16/9/62

lr----- 26 . 35

,.___ _ _ _ 28. 91

llr------ 29.37

..&----- 32. 69

33.34

L-----
38.60

1-o.---.-Jo-.----c:21l-.--~31:-.--:;..,;---_--~~ ·
Figure 8b.
 On teleseismic bodywaves 169
A source was estimated for this event by averaging the log amplitude and phase spectra of
a suite of nine stations which showed the clearest P phases. The seismograms marked with
asterisks in Fig. 8 are the ones which were used in the source estimation. The source wavelet
was windowed by a 20-s window.
Several of the more interesting deconvolutions of the Kern County suite are presented in
the figures following Fig. 8. Each of the seismic traces was delayed so that the first obvious
arrival occurs at the 10 s mark on the arbitrary time scale given with the plots. The impulse
that corresponds to this arrival should appear at the 0 s mark, but to make a clearer pre-
sentation, the impulse response was delayed by 10 s.
The enhancement achieved by deconvolution was judged to be good for this event. The
results are consistent with the work of Heimberger & Wiggins {1971) and Wiggins & Helm-
berger (1973). For every recording, the initial P wave onset is easily discernible, and for the
stations HTMN and MMTN (Figs 12 and 13), the second arrival is very well spiked.
Two problems with this deconvolution method become apparent in these examples. The
first occurs with the waterlevel parameter near unity, i.e. the deconvolution is approaching
a crosscorrelation. For this case, both parts of the source tend to be 'spiked' because the
autocorrelation of the source itself contains three spikes. Thus, it is important to perform
the deconvolution for a number of waterlevels in order that this problem can be detected.
The second problem is over-resolution by the impulse spectrum extension scheme. In
particular, the deconvolution of WNSD may suffer from this problem. It is not clear what is
causing the resonance in this recording. The deconvolution indicates the arrival of two inter-
fering phases, but we believe this result should be treated with suspicion.

5 Conclusions
In this paper we have presented a method for the deconvolution of teleseismic recordings.
The first step in the deconvolution is source estimation. Since earthquakes tend to produce
complex source time functions, the ability to separate the source and transmission path
effects on the basis of a single seismogram is limited. For example, if the earthquake has two
or more successive motions it is difficult, if not impossible, to decide whether the observed
effect on a particular recording is due to the source or the transmission path. Source estima-
tion by homomorphic transforms of single seismograms {Ulrych 1971) was found to be an
unsatisfactory source estimator for earthquakes because of the invalidity of the low que-
frency assumption and the phase instabilities of the transform itself. The obvious remedy for
the source estimation problem is to use the redundant source information available from a
suite of recordings of the same event, whose distance, azimuthal, and time ranges have been
restricted sufficiently to allow the assumption of source uniformity.
The specific method of source estimation that is suggested is to separately ave~age the log
amplitude and phase spectra of the recordings of the suite together. This method is equiva-
lent to averaging the cepstra of the seismograms together. However, it is not necessa.J to
unwrap the phase curve with the result that some of the inherent phase instabilities of the
homomorphic transform are avoided. The low quefrency a3sumption is replaced by the
assumption that averaging sufficiently reduces the cepstral contributions of the impulse
response. This method has the advantage over simple time domain averaging that it uses the
redundant information of secondary arrivals on the recordings. The necessary condition for
this method to resolve the source time function is that the various arrivals on the recordings
have different phase velocities.
The actual deconvolution of the seismograms is accomplished by spectral division. Two
modifications to this method were made. The first is the introduction of a waterlevel para-
meter which constrains the minimum allowable source spectral amplitude level (Heimberger
170 R. W. Gayton and R. A. Wiggins

HKWY p KERN 16/9/62 DELTR=l2.22

0.95
0

0.95
10

0.95
15

0 100

0 100

10

0 1u0

15

0.01
0

0.01
10

0.01
15
D. lu. ';ZD, 31. 4U, :SO.
Figure 9. The deconvolution of HKWY. The format of Fig. 9-14 is the same. The station name, the
primary phase on the record, the name of the event, and the source-receiver separation in degrees are
given in the title. The first trace is the source estimate for the suite, the second trace is the particular
seismogram, and the remaining traces are the various deconvolutions of the seismogram. The decimal
number to the right of each deconvolution is the value of the waterlevel parameter that was used, and the
integer is the autoregressive order that was used in the spectral extension (zero means no extension).
( WNSD p KERN 16/9/62
On teleseismic bodywaves
DELTA=l5o67
171

0095
10

0095
15

0 10
0

0 10
0

10

0 10
0

15

0001
\l
10

0001
XJ~--~~'--~-----------15
Oo 10. 2U. 3lo 40. ~0

Figure 10. The deconvolution of WNSD. See Fig. 9 for a description of the plot format. This is a difficult
seismogram to deconvolve because of the strong resonance effects. The three main spikes in the decon-
volution at high autoregressive order appears to be a case of over-resolution by the spectral extension
scheme.
172 R. W. Gayton and R. A. Wiggins
SEMN P KERN 16/9/62 DELTR=19.81

0.95
0

0.95
10

0.95
15

0. 10
0

0. 10
10

0. 10
15

0.01
0

0.01
10

0.01
15
0. !0. 2J. 3), 4l. 50.

Figure 11. The deconvolution of SEMN, see Fig. 9. The several arrivals that appear in the 10-20 s time
window of the seismogram are well separated in the deconvolution with spectral extension.
 On teleseismic bodywaves 173
HTMN p KERN 16/9/62 DELTR=21.16

0.95
0

0.95
10

0.95
15

0. 10
0

0. 10
10

0. 10
15
i;
!I
0.01
0

0.01
10

0.01
15
o. 10. 2/:J. 30, 4:1. :'"0.

Figure 12. The deconvolution of HTMN, see Fig. 9. The two main arrivals are well resolved in the decon-
volutions. Notice the change in phase between these arrivals.
174 R. W. Gayton and R. A. Wiggins
MMTN p KERN 16/9/62 DELTR=26.3~

0.95
0

0.95
10

0.95
15

0. 10
0

0. 10
10

0. 10
15

0.01
0

0.01
10

0.01
15
0. 10. 2;!. 31, <lD, ~.

Figure 13. The deconvolution of MMTN, see Fig. 9. The two main arrivals are again well separated in the
final deconvolution.
 On teleseismic bodywaves 175
BBPR p KERN 16/9/62 DELTR=32.69

0.95
10

0.95
15

0. 10
10

0. 10
15

0.01
0

0.01
10

0.01
15
D. 10. 31. 3), <lt. !II.

Figure 14. The deconvolution of BBPA, see Fig. 9. The single main arrival is well resolved despite the high
noise level of the seismogram. In the third deconvolution whi::h is actually a crosscorrelation of the source
and the seismogram, the peaks on either side of the main peak do not correspond to arrivals. For this case
both parts of the source are spiked because the autocorrelation of the source itself has three spikes. The
relationship between the envelope and the deconvolution indicates that the average source differs some-
what from the phase of the arrival at 32.7°.
176 R. W. Gayton and R. A. Wiggins
& Wiggins 1971 ). The waterlevel parameter is defined as a fraction of the maximum source
spectral amplitude, and is useful for limiting the gain of the deconvolution in regions where
the seismograms contain little or no information. The parameter also trades off arrival time
resolution with arrival amplitude resolution.
The second modification is the extension of transmission path impulse response beyond
its optimal spectral passband to increase the time domain resolution of the deconvolution
method. This is accomplished by predicting the unknown spectral regions with a prediction
operator determined by Burg's maximum entropy algorithm from the estimated impulse
response spectrum in the optimal passband. The length of the prediction operator and the
waterlevel mentioned above are two parameters which have to be chosen for this method.
Since in practice neither can be chosen optimally, we suggest that the deconvolution be per-
formed for a range of these parameter values. The stability of the deconvolution can be
checked by comparing the results for the various parameter values.
The appearance of a deconvolved seismogram can be substantially improved by super-
imposing the envelope of deconvolution on it (Heimberger & Wiggins 1971 ). This envelope
is independent of the arrival phase shifts and consequently the relationship between the
envelope and the deconvolution indicates the presence of any arrivals phase shifted relative
to the estimated source.

Acknowledgment
This research was supported by the National Research Council of Canada under operating
grant A8854.

References
Akaike, H., 1969. Fitting autoregressive models for prediction, Ann. In st. Stat. Math .• 21, 261-265.
Bogert, B. P., Healy, M. 1. R. & Tukey, 1. W., 1963. The quefrency analysis of time series for echoes:
cepstrum, pseudo-autocovariance, cross-cepstrum, and saphe cracking, Proceedings of the sympo-
sium on time series analysis, pp. 206-243, ed. M. Rosenblatt, Wiley and Sons, Inc., New York.
Bracewell, R., 1965. The Fourier transform and its applications, McGraw-Hill, New York.
Burg, 1. P., 1967. Maximum entropy spectral analysis, 37th Annual International SEG Meeting, Oklahoma
City, 1967 October 31.
Choy, G. L. & Richards, P. G., 1975. Pulse distortion and Hilbert transformation in multiply reflected
and refracted body waves, Bull. seism. Soc. Am., 65, 55-70.
Claerbout, 1. F., 1976. Fundamentals of geophysical data processing, p. 274, McGraw-Hill, New York.
Clayton, R. W., McClary, B. & Wiggins, R. A., 1976. Comments on the paper 'Phase-distortion and Hilbert
transformation in multiply reflected and refracted body waves', by G. L. Choy & P. G. Richards,
Bull. seism. Soc. Am.. 66, 325-327.
Dey Sarkar, S. K., 1974. Upper mantle P-wave velocity distributions beneath Western Canada, PhD thesis,
University of Toronto.
Dey Sarkar, S. K. & Wiggins, R. A., 1976. Source deconvolution of teleseismic P-wave arrivals between
14 o and 40°, J. geophys. Res., in press.
Farnbach, 1. S., 1975. The complex envelope in seismic signal analysis, Bull. seism. Soc. Am., 65,951-
962.
Fryer, G. 1., Odegard, M. E. & Sutton, G. H., 1975. Deconvolution and spectral estimation using final
prediction error, Geophysics, 40,411-425.
Heimberger, D. & Wiggins, R. A., 1971. Upper mantle structure of the Midwestern United States, J.
geophys. Res., 76, 3229-3245.
Hill, D. P., 1974. Phase shift and pulse distortion in body waves due to internal caustics, Bull. seism. Soc.
Am., 64, 1733-1742.
Oppenheim, A. V., 1967. Generalized superposition,IEEE, Information and Control, 11,528-536.
Robinson, E. A., 1967. Statistical communication and detection, Hafner Publishing Co., New York.
 On teleseismic bodywaves 177
Savage, J. C., 1966. Radiation from a realistic model of faulting, Bull. seism. Soc. Am., 56, 577-592.
Schafer, R. W., 1969. Echo removal by discrete generalized linear filtering, Research lab. of Electronics,
MIT, Tech. Report, 466.
Smylie, D. E., Clarke, G. K. C. & Ulrych, T. J., 1973. Analysis of irregularities in the Earth's rotation,
Methods in computational physics, vol. 13, pp. 391-430, Academic Press, New York.
Stoffa, P. L., Buhl, P. & Bryan, G. M., 1974. The application of homomorphic deconvolution to shallow
water marine seismology Part 1: Models, Geophysics, 39,401-416.
Uirych, T J., 1971. Application of homomorphic deconvolution to seismology, Geophysics, 36, 650-660.
Uirych, T. J., Smylie, D. E., Jensen, 0. G. & Clarke, G. K. C., 1973. Predictive filtering and smoothing of
short records by using maximum entropy, J. geophys. Res., 78, 4959-4964.
Uirych, T. J. & Bishop, T. N., 1975. Maximum entropy spectral analysis and autoregressive decomposi-
tion, Rev. Geophys., 13, 183-200.
Wiggins, R. A. & Helm berger, D., 1973. Upper mantle structure of the Western United States, J. geophys.
Res., 78, 1870-1880.