Computational Geosciences 3 (1999) 111–134 111

A wavelet method for detecting S-waves in seismic data

P.J. Oonincx ∗
CWI, P.O. Box 94079, 1090 GB Amsterdam, The Netherlands
E-mail: [email protected]

Received 12 November 1998

Seismic signals consist of several typically short energy bursts, called phases, exhibiting
several patterns in terms of dominant frequency, amplitude and polarisation. We present a
fast algorithm to detect the so-called S-phase in a three-component seismic signal. This new
approach combines traditional S-phase detection methods and the discrete wavelet transform.
Keywords: wavelets, seismology, feature detection, phase pickers, cross-power matrix, en-
ergy distribution
AMS subject classification: 42C15, 65G99, 86A15, 94A12, 94A13

1. Introduction

A seismic earthquake signal recorded by a seismic station (seismogram) is built

up by several different seismic waves (phases), which characterise the type of the
signal. Amongst others, significant phases are the P-phase and the S-phase, which
we consider here. The problem we are dealing with, is to detect automatically the
S-phase and to determine its onset time, once the P-phase arrival time is known with
high accuracy as in [3,14]. An accurate estimate of these arrival times is important
for determining the type and location of the seismic event.
The S-phase arrival time is determined in a three-component seismogram, rep-
resenting motion on a ground detector in three mutually orthogonal directions, two in
the horizontal plane (x–y-plane) and one vertical direction (z-axis). An example of a
three-component seismogram is depicted in figure 1. The detection of the S-phase ar-
rival time is based on some physical differences between the P-phase and the S-phase,
as described thoroughly in [1,4]. For our problem the most obvious property is the
difference in arrival times. The S-phase is always delayed as compared to the P-phase
arrival at the seismic station. A more fundamental property is the fact that P-phases
compress volumes and S-waves deform volumes. Furthermore, the S particle motion,
i.e., the direction of the S-phase when it arrives at the earth’s surface, is contained in a
plane perpendicular to the direction of the P particle motion, called the S-plane. This
property only holds if reflections at the earth’s surface may be ignored or when the
∗
This research is supported financially by the Technology Foundation (STW), project no. CWI44.3403.

 Baltzer Science Publishers BV

112 P.J. Oonincx / Phase detection using wavelets

Figure 1. A three-component seismogram, with P and S arrival times picked by an analyst at iP = 1120
and iS = 1275, respectively.

phases arrive in a direction almost perpendicular to the earth’s surface. In our problem
the latter assumption is justified. The P-phase travels along the travel direction of the
seismic event, unless the medium is anisotropic. Finally, comparing the frequency
spectra of both phases, the P-phase appears at higher frequencies than the S-phase.
The automatic S-phase detection algorithm that we present in this paper is a
combination of traditional methods to detect S-phases as described before and the dis-
crete wavelet transform [5,6,10]. This transform analyses the three components of the
seismogram at several scales (frequency bands). The idea to analyse the components
in this way has already been described in the literature, e.g., [2,11]. However, in these
papers only the wavelet transform itself has been used as a phase detector in seismo-
grams, whereas in our method the wavelet transform has been used in combination
with traditional approaches, which are very well known in seismology and have been
discussed in the past in various papers, e.g., [4,9,13]. Results of this new algorithm
are compared with results based on ideas of Cichowicz [4].
This paper is organized as follows. In section 2 we present some well-known
characteristic functions used in seismology [4,9,13] to detect S-phase arrival times in
a seismogram. In section 3 we deal with the stability of these characteristic functions
 P.J. Oonincx / Phase detection using wavelets 113

with respect to errors in the measured data and an incorrectly determined P-phase
arrival time. Also we consider some problems that appear in the practical use of these
functions. In section 4 we recall [6,10] the concept of multiresolution analysis for
L2 (R) and its relation to the discrete wavelet transform (DWT) on L2 (R). In section 5
we study the use of the DWT for the discrete-time seismic signals we are dealing with.
Furthermore, the wavelet based algorithm to detect S-phase arrival time is presented.
Examples of its use on two different kinds of seismic events and some first test results
can be found in section 6. Finally, in section 7, we present some conclusions.

2. Characteristic functions for S-phase picking

A common strategy to detect phase arrival times is to construct one or more so-
called characteristic functions. These are discrete-time functions, with some specific
properties at the time sample, where a phase arrives. In this section we discuss some
of them for detecting S-phase arrival times.

2.1. Characteristic functions based on a cross-power matrix

We consider a (real-valued) three-component seismic signal u ∈ l2 (Z, R3 ). The

time-dependent N -point cross-power matrix is then defined as follows.

Definition 2.1. Let N ∈ N and u1 , u2 , u3 ∈ l2 (Z). Then the N -point cross-power

matrix of u1 , u2 and u3 at i ∈ Z is given by
 
hu1 , u1 ii hu1 , u2 ii hu1 , u3 ii

MN ,u (i) = hu2 , u1 ii hu2 , u2 ii hu2 , u3 ii  , (2.1)
hu3 , u1 ii hu3 , u2 ii hu3 , u3 ii
for i ∈ Z. In this definition hun , um ii , i ∈ Z, denotes the semi-inner product given by
i+N −1
1 X
hun , um ii = un (k) · um (k).
N
k=i

Note that h·, ·ii depends also on N . The window length N itself depends on the sample
rate and the kind of seismic event.

In practice, we first create a signal u, with zero mean at each component, by

subtracting the means of the components from the seismic signal. The reason for
doing this is to neglect possible offsets without seismic cause. These can be generated
by the measurement equipment.
We observe that MN ,u (i) is the Gram matrix of the set {χ[i,i+N −1] u1 ,
χ[i,i+N −1] u2 , χ[i,i+N −1] u3 }, with χX the characteristic function on Z of a discrete
set X. So MN ,u (i) is a positive semi-definite matrix. Therefore the eigenvalues of
MN ,u (i) are real and positive, λ1 (i) > λ2 (i) > λ3 (i) > 0, and the eigenvectors can be
114 P.J. Oonincx / Phase detection using wavelets

chosen to form an orthonormal basis in R3 . We introduce two characteristic functions

based on the eigenvalues and eigenvectors of MN ,u . Hereby we assume MN ,u (i) 6= 0,
for any i ∈ Z, which is quite a realistic assumption.

2.1.1. Deflection angle

Let v 1 (i) denote the eigenvector of MN ,u (i) corresponding to λ1 (i). This eigen-
vector v 1 (i) represents the direction of the particle motion at time i with most seismic
energy. Let iP be the P-phase arrival time. Then v 1 (iP ) is the direction of the P particle
motion. The deflection angle is defined by
 
2 |(v 1 (i), v 1 (iP ))|
κ1 (i) = arccos . (2.2)
π kv 1 (i)k · kv 1 (iP )k

Note that κ1 (iP ) = 0. Furthermore, since the direction of the S-phase particle motion
is perpendicular to v 1 (iP ), κ1 attains its maximum 1 at the S arrival time iS .

2.1.2. Degree of polarisation

Following [13], the degree of polarisation is defined by

(λ1 (i) − λ2 (i))2 + (λ1 (i) − λ3 (i))2 + (λ2 (i) − λ3 (i))2

κ2 (i) = . (2.3)
2 · (λ1 (i) + λ2 (i) + λ3 (i))2

This characteristic function can be used both for detecting P-phase arrival times and
S-phase arrival times, since all types of seismic polarisation, i.e., all seismic energy
that can be found in one direction, exhibit a high degree of polarisation. This means
that we may expect maxima for κ2 at both iP and iS , the S-phase arrival time.
Note that κ1 and κ2 depend on MN ,u , and thus they depend on the window length
N . Later in this paper we will also discuss window lengths that depend on the kind
of analysed signal. Also we observe that for all characteristic functions κ introduced
in this section we have
∀i ∈ Z 0 6 κ(i) 6 1. (2.4)
A combination of three characteristic functions κ1 , κ2 , κ3 will be used such that

Y
3 Y
3
κ2m (i) = max κ2m (n) ⇐⇒ i = iS . (2.5)
n∈Z
m=1 m=1

We use the square product of the characteristic functions to emphasise the maximal
values attained in each function at iS and to reduce other (local) maxima, related
to features in the signal other than the S-phase arrival. We have already met two
candidates to appear in this product of functions. In the next section we introduce the
third function that can be used in (2.5).
 P.J. Oonincx / Phase detection using wavelets 115

2.2. Rotation and the energy ratio

We transform the three-component seismic signal into a three-component signal

with components in the longitudinal direction and two transversal directions. The lon-
gitudinal direction is the direction of the P particle motion. The transversal directions
are mutually orthogonal and are chosen in the plane perpendicular to the longitudinal
direction. This transversal plan is also called the S-plane, since the direction of the
S particle motion is in the S-plane.
For the P-phase arrival time iP , we compute the eigenvalues λ1 (iP ) > λ2 (iP ) >
λ3 (iP ) > 0 of MN ,u (iP ) and their corresponding mutually orthonormal eigenvectors
v 1 (iP ), v 2 (iP ) and v 3 (iP ). The eigenvalues indicate the distribution of seismic energy
along three directions, given by their corresponding eigenvectors. Now, the direction
corresponding to the largest amount of seismic energy is given by v 1 (iP ). This vector
corresponds to the longitudinal direction. Orthogonal to v 1 (iP ) we have the S-plane
spanned by v 2 (iP ) and v 3 (iP ). In this plane most of the energy can be detected at iS .
The seismic signal is transformed into the basis {v 1 (iP ), v 2 (iP ), v 3 (iP )} by
   
uL (i) u1 (i)
 uQ (i)  = V (iP )  u2 (i)  , (2.6)
uT (i) u3 (i)
with

T
V (iP ) = v 1 (iP ) | v 2 (iP ) | v 3 (iP ) . (2.7)
The third characteristic function we use is the fraction of energy in the S-plane
to the total amount of energy in the signal, given by
Pi+N −1
n=i (uQ (n)2 + uT (n)2 )
κ3 (i) = Pi+N −1 , (2.8)
n=i (uL (n)2 + uQ (n)2 + uT (n)2 )
for some N ∈ N. Note that we may expect a minimum κ3 (iP ) = 0 and a maximum
κ3 (iS ) = 1. Further, κ3 satisfies (2.4).

3. On the use of characteristic functions

In this section we discuss the stability of the characteristic functions that we

considered in the previous section. Further, we discuss problems one has to deal with
when using the characteristic function in practice.

3.1. Error analysis of the characteristic functions

We consider three kinds of error that can appear in the characteristic functions κ.
Amongst others an incorrect κ can be the result of computational and truncation errors
in the matrices MN ,u , measurement errors in the signal u and the determination of an
116 P.J. Oonincx / Phase detection using wavelets

incorrect P-phase arrival time i0P . We will show that κ depends continuously on the
errors as described above.
First we consider stability with respect to computational and truncation errors in
MN ,u and measurement errors in u.
By definition it is clear that κ3 is stable with respect to errors in u and that it
is not affected by errors in MN ,u , since it does not depend on MN ,u . To prove the
stability of κ1 and κ2 we present some auxiliary results from linear algebra.

Theorem 3.1. Let M = A + ∆, with M , A and ∆ (n × n) matrices with k∆k2  1.

Further, let α1 , . . . , αn be the eigenvalues of A and let u1 , . . . , un be their correspond-
ing eigenvectors. Finally, we assume that all eigenvalues have multiplicity 1. Then a
first order approximation of the eigenvectors v 1 , . . . , v n of M is given by
X (∆ui , wl )
v i = ui + u, (3.1)
(αi − αl )(ul , wl ) l
l6=i

with w1 , . . . , wn the eigenvectors corresponding to the eigenvalues α1 , . . . , αn of A∗ .

Proof. Cf. [8]. 

With this theorem we can prove the following corollary.

Corollary 3.2. Let H be a Hermitian (n × n) matrix, n > 1, with eigenvectors

u1 , . . . , un , kuk k = 1, k = 1, . . . , n. Further we assume that all eigenvalues have
multiplicity 1. Then the function gk (H) = uk , k = 1, . . . , n, is continuous.

Proof. Let 0 < ε  1 and let ∆ be an (n×n) matrix, n > 1, with k∆k2 < ε·α/(n−1)
with α = minl6=k |αk − αl |. Here α1 , . . . , αn are the eigenvalues of H. Then, by (3.1),
X (∆u , u ) X (∆u , u )
gk (H + ∆) − gk (H) = k l
u 6 k l
2 αk − αl l 2 αk − αl
l6=k l6=k
6 (n − 1)k∆k2 /α < ε. 

Since κ1 can be written as in corollary 3.2, errors in κ1 depend continuously on

errors in MN ,u (i) under the condition that |σ(MN ,u (i))| = 3, with σ(A) denoting the
spectrum of some matrix A ∈ Rn×n . From [8] we take the following result to prove
the stability of κ2 .

Theorem 3.3. Let M = A + ∆, with M , A and ∆ Hermitian (n × n) matrices. Let

further µ1 > · · · > µn , α1 > · · · > αn and δ1 > · · · > δn be the eigenvalues of M , A
and ∆, respectively. Then
αk + δn 6 µk 6 αk + δ1 . (3.2)
 P.J. Oonincx / Phase detection using wavelets 117

Using this theorem we arrive at the following corollary.

Corollary 3.4. Let Hn denote the set of all Hermitian (n×n) matrices and let λ : Hn →
Rn be given by λ(H) = (λ1 , . . . , λn ), with λ1 > · · · > λn > 0 the eigenvalues of
H ∈ Hn . Then λ is continuous on Hn .

Proof. Let ε > 0 and let ∆ be a Hermitian (n × n) matrix, with eigenvalues δ1 >
· · · > δn and such that k∆k2 < ε. Further, let λ01 > · · · > λ0n denote the eigenvalues
of H + ∆. Then by (3.2)

λ0k − λk 6 max |δ1 |, |δn | 6 k∆k2 < ε.
Therefore kλ(H + ∆) − λ(H)k2 < ε. 

Corollary 3.4 yields immediately

Corollary 3.5. Let f : Rn → [0, 1] be continuous and let Hn denote the set of all
Hermitian (n × n) matrices. Then κ = f ◦ λ is continuous on Hn , with λ as defined
in corollary 3.4.

Since κ2 can be written as in corollary 3.5, also errors in κ2 depend continuously

on errors in MN ,u . Under the restriction that |σ(MN ,u (i))| = 3, we have now proved
0
∀ε > 0 ∃δ > 0 MN ,u (i) − MN ,u (i) 2
< δ =⇒ κn (i) − κ0n (i) < ε, n = 1, 2,
(3.3)
for any fixed i ∈ Z and with κ0n the characteristic function associated with the perturbed
matrices MN 0 . We will show that matrices M
,u N ,u depend continuously on u in the
following sense:

Lemma 3.6. Let u, w ∈ l2 (Z, R3 ) and let MN ,u and MN ,w be the cross-power matrices
associated with u and w, respectively. Then
∀ε > 0 ∃δ > 0 ∀i ∈ Z ku − wk∞ < δ =⇒ MN ,u (i) − MN ,w (i) 2
< ε. (3.4)

Proof. Let d ∈ l2 (Z, R3 ) be defined as d(n) = u(n) − w(n) and let ∆(i) ∈ R3×3 be
given by ∆(i) = MN ,u (i) − MN ,w (i). With a straightforward calculation we get
∆k,l (i) = huk , ul ii − hwk , wl ii = huk , ul ii − huk − dk , ul − dl ii
= hdk , ul ii + huk , dl ii − hdk , dl ii 6 2N kdk∞ · kuk∞ + N kdk2∞ .
So k∆(i)k∞ → 0 if ku − wk∞ → 0. The proof is completed by the equivalence of
matrix norms. 

Resuming, for all characteristic functions of section 2 we proved stability with

respect to computational and truncation errors in MN ,u and measurement errors in u.
118 P.J. Oonincx / Phase detection using wavelets

The last kind of error in the characteristic functions we discuss here is the error
due to an incorrectly determined P-phase arrival time i0P .
With a straightforward calculation we get MN ,V u (i) = V MN ,u (i)V T for all
(3 × 3) matrices V . For orthonormal matrices V this relation yields σ(MN ,V u (i)) =
σ(MN ,u (i)). In particular we have σ(MN ,V (iP )u (i)) = σ(MN ,V (i0P )u (i)) yielding that
κ2 is invariant under any orthonormal transformation of u. So an incorrect P-phase
arrival time i0P will not affect κ2 .
The deflection angle κ1 is affected by an incorrect iP . To get some insight how
κ1 is affected, we rewrite





κ1 (i) = 2 arccos v1 (i), v 1 (iP ) /π = 2 arccos v1 (i), V (iP )T e1 /π



= 2 arccos V (iP )v 1 (i) 1 /π. (3.5)

An expression for the error in κ1 due to an incorrect iP is given in the following

lemma.

Lemma 3.7. Let κ01 (i) be given by 2 arccos(|(V (i0P )v 1 (i))1 |)/π and let δ(i) be defined
as δ(i) = |(V (i0P )v 1 (i))1 | − |(V (iP )v 1 (i))1 |. Then
p

κ1 (i) − κ01 (i) = 2 arccos x(i) + (1 − x(i))2 − δ(i)2 /π, (3.6)

with x(i) = |(V (iP )v 1 (i))1 |2 + δ(i)|(V (iP )v 1 (i))1 |, for all i ∈ Z.

Proof. Using a trigonometric formula we derive

p p

| arccos a − arccos b| = arccos ab + 1 − a2 1 − b2
p p

= arccos a2 + ad + 1 − a2 1 − (a + d)2
p

= arccos a2 + ad + (1 − a2 − ad)2 − d2 ,

with d = b − a. The proof is completed by substituting a = |(V (iP )v 1 (i))1 | and

d = δ(i). 

Using (3.6) we can prove the following lemma that gives conditions on the
signal u, such that the deviation in κ1 due to an incorrectly determined P-phase arrival
time i0P is small.

Lemma 3.8. Let u ∈ l2 (Z, R3 ) and let R denote the shift on l2 (Z, R3 ) given by
(Ry)(k) = y(k − 1). Further, let MN ,u be the cross-power matrix associated with u
and assume |σ(MN ,u (iP ))| = 3. Then
0
∀ε > 0 ∃δ > 0 u − RiP −iP u ∞
< δ =⇒ κ1 − κ01 ∞
< ε, (3.7)

with κ1 and κ01 as given in lemma 3.7.

 P.J. Oonincx / Phase detection using wavelets 119

0 0
Proof. If ku − RiP −iP uk∞ → 0, lemma 3.6 yields with w = RiP −iP u,
MN ,u (iP ) − MN ,u (i0P ) 2
→ 0.
Now, by corollary 3.2 and the assumption |σ(MN ,u (iP ))| = 3 we get
v k (i0P ) − v k (iP ) 2
→0

for k = 1, 2, 3. This yields kV (i0P )T − V (iP )T k2 → 0. Taking δ(i) as in lemma 3.7, we

derive





δ(i) = V (i0P )v 1 (i) 1 − V (iP )v 1 (i) 1 6 V (i0P ) − V (iP ) v 1 (i) 1

T

= v1 (i), V (i0P ) − V (iP ) e1 6 V (i0P )T − V (iP )T 2
→ 0,
for all i ∈ Z. So by (3.6) we get
p

κ1 (i) − κ01 (i) = 2 arccos x(i) + (1 − x(i))2 − δ(i)2 /π
p

→ 2 arccos x(i) + (1 − x(i))2 /π = 2 arccos(1)/π = 0,
which completes the proof, since this result holds for all i ∈ Z. 

We conclude from lemma 3.8 that not only i0P − iP has to be small in order
to get small errors in κ1 , but also u has to be a signal of bounded variation and
|σ(MN ,u (iP ))| = 3.
Finally, we derive an expression for the error in κ3 due to an incorrectly deter-
mined P-phase arrival time.

Lemma 3.9. Let κ03 be as defined in (2.8) with u0L , u0Q and u0T substituted for uL , uQ
and uT respectively. Further we define δ = v 1 (iP ) − v 1 (i0P ). Then

κ3 − κ03 ∞
6 δ v 1 (iP )T + v 1 (iP ) δ T + δ δ T 2
, (3.8)

with k · k2 the operator norm on R3 .

Proof. After rewriting

,
X−1
i+N X−1

i+N
T
κ3 (i) = V (iP ) QV (iP )u, u (u, u)
n=i n=i

with Q the orthonormal projection onto span {e2 , e3 }, we derive

κ3 (i) − κ03 (i)
, i+N −1
X−1
i+N

T

X

= V (iP )T QV (iP ) − V i0P QV i0P u(n), u(n) u(n), u(n)
n=i n=i
120 P.J. Oonincx / Phase detection using wavelets

T

6 V (iP )T QV (iP ) − V i0P QV i0P 2
= V (iP )T P V (iP ) − V (i0P )T P V (i0P ) 2
= v 1 (iP )v 1 (iP )T − v 1 (i0P )v 1 (i0P )T 2
,
using P = I − Q. Since this upper bound holds for any i ∈ Z, we can take the
supremum of |κ3 (i) − κ03 (i)| over all i ∈ Z. The proof is completed by substituting
v 1 (i0P ) = v 1 (iP ) − δ. 

We observe that the upper bound in (3.8) is sharp, so that it is a good indication
for the error in the ratio of transversal to total energy, due to an incorrect P arrival
time determination.

3.2. Problems in analysing seismic data

When analysing a three-component signal with κ = κ1 · κ2 · κ3 we have to deal

with several problems. First of all, the window length N for the matrices MN ,u and
in κ3 has to be determined. Obviously, N is related to the frequency spectrum of the
signal. By fixing N we do not take into account that a signal can consist of a broad
range of frequency contents. Moreover, the window lengths N that are not related to
the frequency contents of the signal will introduce undesired spikes in the graph of κ,
which has been depicted in figure 2. In this figure we see the seismic signal of figure 1

Figure 2. The characteristic function κ for the seismic signal in figure 1 (iP = 1120, iS = 1275) for
N = 15 (top) and N = 30 (bottom).
 P.J. Oonincx / Phase detection using wavelets 121

(iP = 1120, iS = 1275) analysed by κ using two different window lengths, namely
N = 15 and N = 30. Due to an incorrect window length we see in the upper graph
a lot of local maxima that are not related to any specific features in the signal.
Another problem we have to deal with is the following. Ideally κ should attain its
global maximum at iS . Moreover, this maximal value should be close to 1. However,
in practice a seismic signal does not only consist of an P-phase and an S-phase,
but also of other waves, which we did not consider here, appear in the seismogram.
Furthermore, the signal is generally measured with both background noise and signal
generated noise. Due to these facts κ will generally not reach a value close to 1 at iS
and even the maximum of κ at iS can turn out to be a local maximum instead of a
global maximum. Therefore we take a threshold value that has to be attained by κ at
the S-phase arrival time iS . In practice it turns out that choosing this threshold value
is very difficult. To illustrate this phenomenon we may have another look at figure 2.
In both pictures we notice that the global maximum of κ is much less than its ideal
value 1. Furthermore, we see that in the second picture κ attains its global maximum
at i = 1540 while iS = 1290.

4. The discrete wavelet transform

We introduce the discrete wavelet transform (DWT) as a signal processing tool.

By means of filters associated with the DWT we are able to decompose seismic signals
u in time at different scales. With this decomposition we expect that it is possible
to separate the S-phase from the P-phase and probably also from other phases in the
signal.

4.1. Multiresolution analysis

Before introducing the concept of a multiresolution analysis (MRA) in L2 (R),

we define two orthonormal operators.

√ For all f ∈ L (R) the dilation operator

Definition 4.1. 2 D2 on L2 (R) is given by
2
(D2 f )(x) = 2f (2x), and the shift operator T on L (R) is given by (T f )(x) =
f (x − 1).

With these two operators we come to the definition of an MRA, following [5,6,
10].

Definition 4.2. An MRA in L2 (R) is an increasing sequence of closed subspaces

Vj , j ∈ Z, in L2 (R),
· · · ⊂ V−2 ⊂ V−1 ⊂ V0 ⊂ V1 ⊂ V2 ⊂ · · · ,
such that
S
1. j∈Z Vj is dense in L2 (R),
122 P.J. Oonincx / Phase detection using wavelets

T
2. j∈Z Vj = {0},

3. f ∈ Vj ⇐⇒ D2 f ∈ Vj+1 , ∀j ∈ Z,

4. ∃φ ∈ L2 (R): {T k φ | k ∈ Z} is an orthonormal basis for V0 .

A more general concept of MRA and its application to the seismic problem we
are dealing with can be found in [12]. There we replaced condition 4 of definition 4.2
by
 k
∃φ ∈ L2 (R): T φ | k ∈ Z is a Riesz basis for V0 .
Characterisations of a Riesz basis are given in [16].
By the definition of an MRA an orthonormal basis for Vj , with j ∈ Z fixed, is
given by {D2j T k φ | k ∈ Z}, once such a function φ has been found. Such φ is called
a scaling function. Since D2 is an orthonormal operator on L2 (R) and V0 is invariant
under the action of D2 , the collection {D2 T k φ | k ∈ Z} is an orthonormal basis for
V1 . As we also have φ ∈ V1 , we get
X
φ= h(k)D2 T k φ, (4.1)
k∈Z

for some h ∈ l2 (Z). Later on in this paper we want h to generate a bounded convolution
operator on l2 (Z). Therefore we require h ∈ l1 (Z). We will refer to relation (4.1) as
a scale relation and to h as a scale sequence.
We consider again the inclusion V0 ⊂ V1 . Obviously we can define a subspace W0
such that W0 ' V1 /V0 . To come to a unique definition of W0 , we take W0 = V1 ∩ V0⊥ .
Using the invariance of the subspaces Vj under the action of the orthonormal operator
D2 we arrive in a natural way at the definition of the closed subspaces Wj ⊂ L2 (R)
by putting Wj = Vj+1 ∩ Vj⊥ . Recursively repeating the orthonormal decomposition of
some VJ into VJ−1 and WJ−1 yields
!
M
J−1
VJ = VJ−1 ⊕ WJ−1 = VJ−2 ⊕ WJ−2 ⊕ WJ−1 = · · · = V−J ⊕ Wj .
j=−J

Taking
L J → ∞ and applying conditions 1 and 2 from definition 4.2 leads to
W 2
j∈Z j = L (R).
We can find a function ψ ∈ V1 such that {T k ψ | k ∈ Z} is an orthonormal basis
for W0 . Then {Dj T k ψ | k ∈ Z} is an orthonormal basis for Wj , j ∈ Z. Since the
subspaces Wj are chosen to be mutually orthogonal, we then have an orthonormal
basis in L2 (R) given by {Dj T k ψ | j, k ∈ Z}. The function ψ is called a wavelet and
{Dj T k ψ | k ∈ Z} a wavelet basis for Wj , for fixed j ∈ Z. So, using these wavelet
bases we are able to decompose any f ∈ L2 (R) into functions at several scales.
 P.J. Oonincx / Phase detection using wavelets 123

Since the wavelet function ψ should be in V1 , there exists also a scale relation
for ψ:
X
ψ= g(k)D2 T k φ, (4.2)
k∈Z

for some g ∈ l2 (Z), the scaling sequence for ψ. As in (4.1) we will also require
g ∈ l1 (Z). By this relation, the problem of finding ψ can be substituted by the
problem of finding g if φ, and therefore also h, is known. A well-known choice [6]
for g is given by
g(k) = (−1)k h(1 − k). (4.3)

4.2. Decomposition and reconstruction

Let Pj and Qj denote the orthoprojectors on Vj and Wj , respectively. By defin-

ition we have
Pj+1 = Pj + Qj , (4.4)
for all j ∈ Z. We discuss a fast method to decompose Pj+1 f into Pj f and Qj f and
to recover Pj+1 f from Pj f and Qj f .

4.2.1. The decomposition algorithm

We assume Pj+1 f ∈ Vj+1 is known for a certain j ∈ Z. So there exists a
sequence cj+1 ∈ l2 (Z) such that
X
Pj+1 f = cj+1 (k)D2j+1 T k φ.
k∈Z

Moreover, the sequence cj+1 is given by cj+1 (k) = (D2j+1T k φ, f )2 . Following (4.4)
we have
X X
Pj+1 f = cj (k)D2j T k φ + dj (k)D2j T k ψ. (4.5)
k∈Z k∈Z

The sequences cj and dj can be calculated respectively by cj (k) = (D2j T k φ, f )2 and

dj (k) = (D2j T k ψ, f )2 . However, there is also a fast method to derive cj and dj from
the known sequence cj+1 without calculating inner products.
For cj we derive, using (4.1) and (4.5),

X

cj (k) = Pj+1 f , D2j T k φ 2 = h(n) Pj+1 f , D2j+1T 2k+n φ 2
n∈Z
X

= cj+1 (m)h(n) D2j+1T m φ, D2j+1T 2k+n φ 2
m,n∈Z
X


= cj+1 (2k + n)h(n) = W2 cj+1 ∗ ȟ (k), (4.6)
n∈Z
124 P.J. Oonincx / Phase detection using wavelets

with ȟ(n) = h(−n) and with W2 the downsampling operator given by

(W2 y)(k) = y(2k),
for all y ∈ l2 (Z). In the same manner we get
dj = W2 (cj+1 ∗ ǧ). (4.7)
Recursively we get expressions for cj−n and dj−n for n > 1, namely
cj−n = (W2 Cȟ )n cj and dj−n = W2 Cǧ (W2 Cȟ )n−1 cj , (4.8)
where Cx denotes convolution with x ∈ l1 (Z).

4.2.2. The reconstruction algorithm

Once we have computed a decomposition of Pj+1 f into Pj f and Qj f by means
of the coefficients cj and dj , we can also recover cj+1 out of cj and dj in an efficient
way. In order to derive such a reconstruction formula we will represent Pj f and Qj f
in terms of D2j+1T k φ, k ∈ Z, the basis functions of Vj+1 .
Since Pj f ∈ Vj+1, there exists an αj ∈ l2 (Z) such that
X
Pj f = αj (k)D2j+1 T k φ.
k∈Z

Taking the inner product with D2j+1 T k φ in both the left hand side and the right hand
side of this equation yields, in combination with (4.1),

X

αj (k) = Pj f , D2j+1T k φ 2 = cj (n) D2j T n φ, D2j+1T k φ̃ 2
n∈Z
X j+1 m+2n

= cj (n)h(m) D2 T φ, D2j+1 T k φ̃ 2
m,n∈Z
X

= cj (n)h(k − 2n) = (U2 cj ) ∗ h (k), (4.9)
n∈Z

with U2 the upsampling operator given by


y(k/2), if k mod 2 = 0,
(U2 y)(k) =
0, otherwise,

for all y ∈ l2 (Z). We observe that W2 U2 = I and (W2 )∗ = U2 . In the same manner
we get
X
Qj f = βj (k)D2j+1 T k φ,
k∈Z

with
βj = (U2 dj ) ∗ g. (4.10)
 P.J. Oonincx / Phase detection using wavelets 125

So we derived the reconstruction formula

cj+1 = αj + βj = (U2 cj ) ∗ h + (U2 dj ) ∗ g.
Recursively we get expressions for αj−n and βj−n for n > 1, namely
αj−n = (Ch U2 )n (W2 Cȟ )n cj , (4.11)
βj−n = (Ch U2 ) n−1
Cg U2 W2 Cǧ (W2 Cȟ )
n−1
cj . (4.12)
We observe that by (4.11) and (4.12) coarse scale information is represented at the
finest resolution. In this manner problems with shift variance of the downsampled
DWT are overcome. In the next section we will use these expressions to introduce a
discrete wavelet transform (DWT) for l2 (Z).

5. The DWT for seismic analysis

We discuss how the introduced DWT can be used in combination with the char-
acteristic functions of section 2 to analyse three-component seismograms. Since the
seismic data are in l2 (Z) we first consider the relation between the presented DWT and
the analysis of sampled signals at scales. After that we discuss how the characteristic
function κ can be used in an efficient way to analyse seismic data at several scales.

5.1. Analysing discrete-time signals with the DWT

In section 4 we introduced the DWT for functions s ∈ L2 (R). In order to analyse

signals in l2 (Z) at several scales we introduce a decomposition at scales for sequences
in l2 (Z). This is done by identifying s ∈ l2 (Z) with some wavelet coefficients cj .

Definition 5.1. Let s ∈ l2 (Z). Then the l2 -DWT of s at scale m is defined as

s(m) = (Ch U2 )m−1 Cg U2 W2 Cǧ (W2 Cȟ )m−1 s, (5.1)
for m ∈ Z. Here h and g denote the scale sequences as given in (4.1) and (4.2).

From now on we will use the operators A = W2 Cȟ and B = W2 Cǧ . Since
we required g, h ∈ l1 (Z), A and B are bounded operators with kAk2 = khk1 and
kBk2 = kgk1 . Using these operators we have s(m) = (A∗ )m−1 B ∗ BAm−1s. We
observe that (4.12) can also be seen as the l2 -DWT of cj+1 at scale j + 1 − m, i.e.,
βm = c(j+1−m)
j .
Besides the l2 -DWT of a sequence s ∈ l2 (Z), we also define its approximation
∗ m m (j+1−m)
sequence s(m)
a a = (A ) A s. Note that αm = (cj )a
at level m by s(m) .
The l2 -DWT we proposed in this paper is a special case of the DWT for l2 (Z)
that Cohen proposed in [5] using a dyadic dilation operator on l2 (Z) and the shift
operator on l2 (Z) to construct an MRA for l2 (Z).
126 P.J. Oonincx / Phase detection using wavelets

Following the recursive approach of section 4.2 we get

X
M
s= s(M
a
)
+ s(m) ,
m=1

for all M ∈ N, by substituting s = cj+1 . Following the orthogonal construction of the

subspaces Vj and Wj we can show rather easily that sequences s(M a
) and s(m) , m =

1, . . . , M , are mutually orthonormal if {R g | k ∈ Z} and {R h | k ∈ Z} are

2k 2k

orthonormal systems in l2 (Z), with also (g, R2k h)2 = 0 for all k ∈ Z. Here R denotes
the shift operator on l2 (Z) given by (Ry)(k) = y(k − 1). As a corollary we get

2
X
M
2
ksk22 = s(M
a
)
2
+ s(m) 2
, (5.2)
m=1
saying that the energy in s is equal to the energy of its decomposition signals at each
scale using the l2 -DWT plus its approximation at a certain level. This result can also
be related to the wavelet function ψ and its scaling function φ. In [15] we proved that
{T k φ | k ∈ Z} and {T k ψ | k ∈ Z} are orthonormal bases for V0 and W0 , respectively,
if and only if {R2k g | k ∈ Z} ∪ {R2k h | k ∈ Z} is an orthonormal basis for l2 (Z).
Combining this result with (5.2) yields

Lemma 5.2. Let {T k φ | k ∈ Z} and {T k ψ | k ∈ Z} be orthonormal bases for V0 and

W0 , respectively. Then
X
M
) 2 2
ksk22 = s(M
a 2
+ s(m) 2
m=1
and conversely.

a k2 > ksa
Also from (5.2) we derive in a straightforward way ks(M ) (M +1) k > 0.
2
This yields that limm→∞ ks(m) k2 exists. In the following theorem we show that
limm→∞ ks(m) k2 = 0.

Theorem 5.3. Let s ∈ l2 (Z) and h ∈ l1 (Z), such that {R2k h | k ∈ Z} is an orthonor-
mal system in l2 (Z). Let, further, s(m)
a be given by s(m)
a = (A∗ )m Am s for m ∈ N.
Then
s(m)
a 2
→0 (m → ∞). (5.3)

Proof. By the orthonormality of R2k h, k ∈ Z, we get AA∗ = I and thus

m
m


= A∗ Am s, A∗ Am s 2 = Am s, Am s 2 = Am s 2 .
2 2
s(m)
a 2
In [5] Cohen proved that kAm δk k2 tends to zero as m → ∞, with {δk | k ∈ Z} the
standard orthonormal basis in l2 (Z). So ks(m)
a k2 → 0 (m → ∞) for any s ∈ l0 (Z) =
2

span{δk | k ∈ Z}. Since l02 (Z) is dense in l2 (Z) the result holds for any s ∈ l2 (Z). 
 P.J. Oonincx / Phase detection using wavelets 127

So, if {R2k g | k ∈ Z} ∪ {R2k h | k ∈ Z} is an orthonormal basis of l2 (Z) we can

combine (5.2) and (5.3). This yields
X
M
2
lim s(m) 2
= ksk22 .
M →∞
m=1

Thus ks(m) k2 will tend to zero for m → ∞.

5.2. Characteristic functions and the DWT

The idea to detect S-phase arrival times using the DWT is as follows. By making
a decomposition of the three-component seismic signal u into signals at several scales,
it can be possible to separate the S-phase from other phases that appear at other scales.
This decomposition is made by taking the l2 -DWT of each component of u after
rotation into longitudinal and transverse directions, i.e.,
u(m) = u(m) (m) (m)
L v 1 (iP ) + uQ v 2 (iP ) + uT v 3 (iP ).

It is already known that the frequency spectrum of the P-phase appears at higher
frequencies than the S-phase. Therefore, in general the S-phase will appear at higher
scales than the P-phase. Our aim is now to find the set JS , denoting the collection
of scaling levels at which most of the S-phase appears. To make this choice more
explicit, we compute the energy in the l2 -DWT of a segment w of u just after iP . This
will yield information about JS . Since the measured seismic signal can also consist
of background noise and microseismic noise, we have to subtract the energy of the
noise. We will do this by using a wavelet soft-thresholding method similar to Donoho’s
method [7].
For denoising we assume u = u(s) +u(n) , with u(s) (i) = 0, i < iP , i.e., the seismic
data consist of some noise and a mixture of phases that can be measured after iP . We
construct w+ (i) = u(iP +i), i = 0, . . . , T −1, and w− (i) = u(iP −i), i = 0, . . . , T −1,
with T such that T /fs = 20, for the sampling frequency fs . Further, we take wn+ (i) = 0
and wn− (i) = 0 for 0 < i and i > T . So we compute the l2 -DWT of two segments
with a duration of 20 seconds before and after the P-phase arrival time. Following [7],
the wavelet coefficients BAm−1 wn− of wn− are used as soft thresholds for BAm−1 wn+ .
We define the threshold operator T on l2 (Z) by



(Tη x)(k) = sgn x(k) · max 0, |x(k)| − η , (5.4)

for some η ∈ R. We compute t(m) = kBAm−1 wn− k∞ for m ∈ N and use this as soft
threshold for BAm−1 wn+ . We get the sequences

m−1 ∗
yn(m) = A∗ B Tt(m) BAm−1 wn+ , n = 1, 2, 3, m ∈ N,
representing the denoised segment w+ . Experiments showed that with this denoising
method the sequences yn(m) can become free of microseismic noise.
128 P.J. Oonincx / Phase detection using wavelets

Figure 3. Energy distributions as functions of the scale for (a) a local event, (b) a non-local event.

We compute the energy distribution of y over all scales

X
3
2
E(m) = y (m)
n 2
. (5.5)
n=1

We already observed that the energy distribution will tend to zero once a certain scaling
level has been reached. We can stop computing the l2 -DWT for higher scales at that
moment. In figure 3 the energy distribution E as a function of the scaling level has
been depicted for both a local event, i.e., an event for which the distance from the
source to the measurement equipment is less than 100 km, and a non-local event. In
this analysis the Daubechies-4 (Db4) wavelet filter and its corresponding scaling filter
have been used, see [6].
Experiments showed that for local events the S-phase arrival time can be noticed
within this time period. For non-local events the difference iS − iP will be larger than
20 seconds in general, due to the fact that for these events the traveling distance of
the phases, i.e., the distance between the seismic station and the source of the event, is
much larger. Another difference between local and non-local events is the frequency
behaviour of the phases. Both the P-phase and the S-phase for local events are high
frequency signals compared to the P-phase and S-phase of non-local events. Therefore
most of the energy of w is found at low scaling levels for local events and at higher
 P.J. Oonincx / Phase detection using wavelets 129

scales for non-local events. For local events iS is contained in w+ , however, in general
iS − iP > T for non-local events. So only iP is contained in w+ for non-local events.
To analyse the energy distribution we consider the scale mmax for which the
maximum of E is attained. In case of a local event this maximum will be related to
the S-phase, which contains most of the energy in y. For non-local events mmax can
be related to the P-phase. Let us now assume that we know that most of the energy
of local events can be found generally at the first mloc scales. Then in general an
event can be characterized as a local event if mmax 6 mloc and as a non-local event if
mmax > mloc . The parameter mloc is determined experimentally and depends on the
wavelet filters. Experiments with the Daubechies-4 wavelets yielded mloc = 4.
Now that we know which event we are dealing with, we take

{mmax , mmax + 1}, if mmax 6 mloc ,
JS = (5.6)
{mmax + 1, mmax + 2}, if mmax > mloc .

The latter choice is justified by observing that the S-phase will appear at higher scales
than the P-phase. Experiments showed that these choices for JS only lead to scales at
which most of the S-phase appears for signals free of microseismic noise. Microseismic
noise with much energy may appear at other scales than the S-phase.
Once JS has been determined κ can be applied to u(m) for m ∈ JS and the
S-phase arrival time is then given by
Y

2 Y

2
κ u(m) (i) = max κ u(m) (n) ⇐⇒ i = iS . (5.7)
n∈N
m∈JS m∈JS

In practice we have to work with a threshold value that has to be attained to declare
iS . However, by separating the S-phase from the P-phase it will be less difficult to
find an appropriate threshold value than in the situation we discussed in section 3.2.
Another problem discussed in section 3.2 is the window length N , which was
not related to the frequency behaviour of u in traditional methods [4]. Now that we
use a decomposition of u at several scales, we can use window lengths N (m), i.e., a
monotone ascending function of the scaling level m. So the characteristic function κ
to analyse u(m) will use the window length N (m) in its definition. In this manner the
window length is adapted automatically to the frequency behaviour of the signal. In
the next section we will discuss the choice we made for N (m).

5.3. A set-up for the DWT analysis

In the previous section we discussed our approach to analyse three-component

seismic data using characteristic functions and the DWT. Here we will present some
choices we made for the parameters in our algorithm after testing the method on
seismic data. Note that also other choices for the parameters can be made as long as
they fit in the mathematical framework of our algorithm.
130 P.J. Oonincx / Phase detection using wavelets

5.3.1. Wavelet filters

The first parameter we discuss is the wavelet function and the associated wavelet
filters. We want the wavelet function to match with the seismic data we are dealing
with, so that the dilated and translated wavelets generate a good approximation of the
several phases at the particular scaling levels. In experiments we minimized for a set
of seismic data (also synthetic data) for which iS is known the error in the determined
iS for a collection of candidate orthonormal wavelet filters. The experimental results
led to the choice of the Daubechies-4 (D4 ) wavelet filters [6].
We observe that also other filters associated to wavelets within the Daubechies
family (e.g., D8 , D20 ) performed very well in this test. These wavelets are much
smoother than D4 , see [6], however in our problem smoothness of the wavelet does
not play an important role. The most important property of the chosen wavelet is its
matching property with the seismic signals, which are not smooth at all. Furthermore,
the wavelet filters of D8 and D20 contain respectively two and five times as many
filter coefficients as the D4 wavelet filters. Using more filter coefficients will increase
the computing time of our algorithm.

5.3.2. Window length N (m)

In section 5.2 we already considered a way to adapt the window length N used
in the characteristic functions to the frequency behaviour of the signals. Obviously,
to analyse u at several scales m ∈ N the window length N (m) has to be a monotone
ascending function of the scaling level m. In order to come to an appropriate function
for N (m) one has to consider two facts. The signal is scaled at each level by the
factor 2. Further, for low and high scaling levels N (m) should not become too small
or too big, since N (m) is used to obtain information out of the signal in a certain
neighbourhood. These considerations led to the choice
 
N (m) = 30 · 2max(0,(m−4)/2) , (5.8)

with d·e the entier function. So√for the first four levels we take N = 30 and thereafter
N is multiplied by powers of 2.

5.3.3. Threshold value

To declare an S-phase arrival time we can use (5.7). However, as we discussed
already in practice κ will not attain its maximum value κmax at iS . Therefore we
replace κmax by a threshold value 0 6 γ 6 κmax . Since κmax will vary for a set of
seismic data we also want to make γ adaptive. This can be done by putting

γ = c · κmax . (5.9)

Now the problem is left to choose c. First experiments with 0.6 < c < 0.9 worked out
very well. However, more experiments are needed to come to an efficient value for c.
 P.J. Oonincx / Phase detection using wavelets 131

6. Examples/first results

To illustrate the difference of our approach compared to Cichowicz’ approach,

we have plotted both the characteristic function in (2.5) and the characteristic func-
tion in (5.7) for two events. For both functions we used the set-up as discussed in
section 5.3.

Example 6.1. In figure 4 we see characteristic functions of the local event which
has been depicted in figure 1. Both functions use the window length N = 30 and
after analysing the energy distribution along scales (5.5) we obtained JS = {2, 3}.
Obviously, the parameter c in (5.9) can attain almost any value between 0 and 1
in our approach, while the only possible choices for c in Cichowicz’ approach are
0.55 < c < 0.95. Note, that for automatic phase detection we have to choose a value
for c before analysing a seismic event.
For this event an analyst took iP = 1120 and iS = 1275. For our approach we get
iS = 1280 and for Cichowicz’ approach we have iS = 1295. With a 40 Hz sampling
frequency these results differ by 0.125 and 0.5 seconds, respectively. Both results are
acceptable in an automatic procedure; however, in our approach it was much easier to
determine the S-phase arrival time than in the traditional approach.

Figure 4. The characteristic function κ for the seismic signal in figure 1 (iP = 1120, iS = 1275) using
Cichowicz’ method (top) and the DWT method (bottom).
132 P.J. Oonincx / Phase detection using wavelets

Note that for this event iS − iP = 1275 − 1120 = 155 (3.875 seconds). So
iS − iP < 20 seconds which we assumed for local events.

Example 6.2. In figure 5 the three-component seismogram of a teleseismic event has

been depicted. The characteristic functions of this teleseismic event have been depicted
in figure 6. The first two functions are based on Cichowicz’ method using the window
lengths N = 60 and N = 120, respectively. These values for N correspond to N (6)
and N (8), cf. (5.8). The third function uses our DWT based approach. Computing (5.5)
yielded JS = {8, 9}, and so the product of the characteristic function at scale 8 and
9 has been computed and plotted in figure 6. We see that in the first picture the S-
phase arrival will not be detected unless we put 0.96 < c 6 1. With a larger window
N = 120, the S-phase arrival can be detected if we take 0.55 < c 6 1; however, in our
approach we do not have to be that precise in choosing c, since for all 0.15 < c 6 1
the S-phase arrival will be detected.
An analyst took iP = 1810 and iS = 11900 for this particular event. We get
iS = 11960 for our wavelet approach and iS = 11990 for Cichowicz’ approach. Since
this signal has also been sampled at 40 Hz, these results differ by 1.5 and 2.25 seconds,

Figure 5. A three-component seismogram of a teleseismic event, with P and S arrival times picked by
an analyst at iP = 1810 and iS = 11900, respectively.
 P.J. Oonincx / Phase detection using wavelets 133

Figure 6. The characteristic function κ for the seismic signal in figure 5 (iP = 1810, iS = 11900)
using Cichowicz’ method for N = 60 (top) and N = 120 (middle) and the DWT method (bottom),
respectively.

respectively. Also in this example our approach improves the automatic picking as
compared to the traditional approach.
We observe that for this event iS − iP = 11900 − 1810 = 10090 (25.25 seconds).
So iS − iP > 20 seconds which we assumed for non-local events.

Our method has been implemented at the Royal Dutch Meteorological Institute
(KNMI) and tested for a huge set (319) of local events. Several tests have been
made with different choices for the parameters. Comparing our results with results we
obtained by Cichowicz’ approach showed considerable improvements in determining
iS correctly. These test results shall appear in a forthcoming paper.

7. Conclusions

We presented a new method to detect S-phase arrival times in three-component

seismograms. This method is based on both traditional methods from seismology
based on physical properties of seismic signals and on the (non-downsampled) discrete
wavelet transform. This transform has been implemented using filter operations on the
sampled seismograms.
134 P.J. Oonincx / Phase detection using wavelets

First results show that this method can be very reliable for automatically detecting
S-phases in a seismogram; however, there are still some unsolved questions on how
to pick the optimal values for the several parameters in the algorithm. These first
results are very promising, and convince us that further testing on huge seismic data
sets (of non-local events) will lead to better choices for the parameters involved in our
algorithm.

Acknowledgements

The author likes to thank the KNMI for providing seismic data to test the algo-
rithm. In particular he would like to thank Reinoud Sleeman and Torild Van Eck from
the Seismology Department of the KNMI for their advice and suggestions in the field
of seismology.

References

[1] K. Aki and P.G. Richards, Quantitative Seismology (Freeman, San Francisco, 1980).
[2] K. Anant and F. Dowla, Wavelet transform methods for phase identification in three-component
seismograms, Bull. Seism. Soc. Am. 87 (1997) 1598–1612.
[3] M. Baer and U. Kradolfer, An automatic phase picker for local and teleseismic events, Bull. Seism.
Soc. Am. 77 (1987) 1437–1445.
[4] A. Cichowicz, An automatic S-phase picker, Bull. Seism. Soc. Am. 83 (1993) 180–189.
[5] A. Cohen and R. Ryan, Wavelets and Multiscale Signal Processing (Chapman & Hall, London,
1995).
[6] I. Daubechies, Ten Lectures on Wavelets (SIAM, Philadelphia, PA, 1992).
[7] D.L. Donoho, De-noising by soft-thresholding, Technical Report, Department of Statistics, Stanford
University (1992).
[8] J. Franklin, Matrix Theory (Prentice-Hall, Englewood Cliffs, NJ, 1968).
[9] D.C. Jepsen and B.L.N. Kennett, Three-component analysis of regional seismograms, Bull. Seism.
Soc. Am. 80 (1990) 2032–2052.
[10] S. Mallat, A theory for multiresolution signal decomposition: the wavelet representation, IEEE
Trans. Patern. Anal. Mach. Intell. 11 (1989) 674–693.
[11] G. Olmo and L. Lo Presti, in: Wavelets: Theory, Algorithms and Applications, eds. C. Chui,
L. Montefusco and L. Puccio (Academic Press, London, 1994) pp. 561–572.
[12] P.J. Oonincx, Automatic phase detection in seismic data using the discrete wavelet transform, CWI-
Report PNA-R9811 (1998).
[13] J. Samson and J. Olson, Some comments on the descriptions of the polarization states of waves,
Geophys. J. Roy. Astr. Soc. 61 (1980) 115–129.
[14] R. Sleeman and T. van Eck, Robust automatic P-phase picking: An on-line implementation in the
analysis of broad-band seismogram recordings, to appear in Phys. Earth Planetary Interior.
[15] S.J.L. van Eijndhoven and P.J. Oonincx, Frames, Riesz systems and MRA in Hilbert spaces, CWI-
Report PNA-R9701 (1997).
[16] R.M. Young, An Introduction to Nonharmonic Fourier Series (Academic Press, New York, 1980).