You are on page 1of 22

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/257596714

Comparison of seismicity declustering methods using a probabilistic measure


of clustering

Article  in  Journal of Seismology · July 2013


DOI: 10.1007/s10950-013-9371-6

CITATIONS READS

13 3,158

5 authors, including:

K. Z. Nanjo Kenji Satake

76 PUBLICATIONS   1,136 CITATIONS   
The University of Tokyo
408 PUBLICATIONS   13,762 CITATIONS   
SEE PROFILE
SEE PROFILE

Jiancang Zhuang Hamdache Mohamed


The Institute of Statistical Mathematics Center for Research in Astronomy and Astrophysics Geophysics
139 PUBLICATIONS   2,834 CITATIONS    55 PUBLICATIONS   487 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Tsunami Data Assimilation Without a Dense Observation Network View project

Human Capacity Building at The University of Tokyo towards Knowledge Base Development for Nuclear Risk Management View project

All content following this page was uploaded by Hamdache Mohamed on 01 July 2014.

The user has requested enhancement of the downloaded file.


J Seismol (2013) 17:1041–1061
DOI 10.1007/s10950-013-9371-6

ORIGINAL ARTICLE

Comparison of seismicity declustering methods


using a probabilistic measure of clustering
Abdelhak Talbi & Kazuyoshi Nanjo & Kenji Satake &
Jiancang Zhuang & Mohamed Hamdache

Received: 2 April 2012 / Accepted: 23 April 2013 / Published online: 24 May 2013
# Springer Science+Business Media Dordrecht 2013

Abstract We present a new measure of earthquake background, the obtained clustering measure shows a
clustering and explore its use for comparing the decrease followed by an increase, defining a V-shaped
performance of three different declustering methods. trend, which can be explained by the presence of short-
The advantage of this new clustering measure over and long-range correlation in the inter-event time series.
existing techniques is that it can be used for non- Three previously proposed declustering methods (i.e.,
Poissonian background seismicity and, in particular, to the methods of Gardner and Knopoff, Reasenberg, and
compare the results of declustering algorithms where Zhuang et al.) are used to obtain an approximation of the
different background models are used. We use our ap- residual “background” inter-event time distribution in
proach to study inter-event times between successive order to apply our clustering measure to real seismicity.
earthquakes using earthquake catalog data from Japan The clustering measure is then estimated for different
and southern California. A measure of the extent of values of magnitude cutoffs and time periods, taking
clustering is introduced by comparing the inter-event into account the completeness of each catalog. Plots of
time distributions of the background seismicity to that the clustering measure are presented as clustering atten-
of the whole observed seismicity. Theoretical aspects of uation curves (CACs), showing how the correlation de-
the clustering measure are then discussed with respect to creases when inter-event times increase. The CACs
the Poissonian and Weibull models for the background demonstrate strong clustering at short inter-event time
inter-event time distribution. In the case of a Poissonian ranges and weak clustering at long time ranges. When
the algorithm of Gardner and Knopoff is used, the CACs
show strong correlation with a weak background at the
A. Talbi (*) : K. Nanjo : K. Satake short inter-event time ranges. The fit of the CACs using
Earthquake Research Institute, University of Tokyo, the Poissonian background model is successful at short
1-1-1 Yayoi, Bunkyo-ku, Tokyo 113-0032, Japan and intermediate inter-event time ranges, but deviates
e-mail: abdelhak_t@yahoo.fr
at long ranges. The observed deviation shows that
A. Talbi the residual catalog obtained after declustering remains
e-mail: a.talbi@craag.dz non-Poissonian at long time ranges. The apparent back-
ground fraction can be estimated directly from the CAC
A. Talbi : M. Hamdache
fit. The CACs using the algorithms of Reasenberg and
Département Etude et Surveillance Sismique,
Centre de Recherche en Astronomie Astrophysique Zhuang et al. show a relatively similar behavior, with a
et Géophysique (CRAAG), BP 63 Route de l’observatoire, time correlation decreasing more rapidly than the CACs
Bouzaréah, 16340 Algiers, Algeria of Gardner and Knopoff for shorter time ranges. This
study offers a novel approach for the study of different
J. Zhuang
Institute of Statistical Mathematics, 10-3 Midori-Cho, types of clustering produced as a result of various
Tachikawa, Tokyo 190-8562, Japan hypotheses used to account for different backgrounds.
1042 J Seismol (2013) 17:1041–1061

Keywords Seismicity . Correlations . Clustering . probability law (Kagan and Knopoff 1976; Zhuang et al.
Declustering algorithms . Probability distributions 2002, 2004). The algorithm of Zhuang et al. (2004) uses
the standardized intensity rates of triggered and
background events from the epidemic-type aftershock
1 Introduction sequence (ETAS) intensity rate as probabilities for the
construction of the cluster and background processes,
Seismicity is characterized by a complex pattern of respectively. More recently, non-parametric approaches
events in space and time (e.g., Utsu 2002). In have been considered in order to obtain objective
particular, earthquakes exhibit space–time clustering, declustering algorithms. For example, Hainzl et al.
which is notably enhanced during seismically active (2006) developed a non-parametric reconstruction of
periods following major events or during a series of background rates based on the first- and second-order
related events referred to as “earthquake swarms.” mean proprieties of inter-event times. This method
However, it is still problematic to differentiate between can be used to decluster earthquake catalogs by com-
clustered and background seismicity components due to paring the whole and background probability density
the overlap (superposition) of the two components in functions (van Stiphout et al. 2013). Similarly, Marsan
time and space (e.g., Touati et al. 2009, 2011). and Lengline (2008) used rate weighting to decluster
Currently, there is no standard definition of the dif- earthquake catalogs using assumptions of a linear trig-
ferent series of earthquakes, and a working definition is gering process and a magnitude-dependent mean field
needed to support the use of the terms “aftershock,” response to the occurrence of an earthquake. More
“mainshock,” and “foreshock.” Given this, many recently, Bottiglieri et al. (2009) used the coefficient of
seismological studies have used a series of declustering variation (ratio of the standard deviation to the mean) of
algorithms based on subjective constraints to capture inter-event times to distinguish periods of pure back-
observed clustering (e.g., Utsu 1969; Gardner and ground activity from coeval cluster periods.
Knopoff 1974; Reasenberg 1985; Frohlich and Davis The present study was motivated by the common use
1990; Davis and Frohlich 1991; Knopoff 2000; see van of the mean properties of inter-event times in all these
Stiphout et al. 2012 for a complete review). These aforementioned approaches. The use of inter-event times,
approaches are based on constraints derived from instead of occurrence rates, is motivated by the stability
characteristic space–time patterns of seismicity follow- of the statistics based on the mean inter-event time, t
ing major events (Omori 1894; Utsu 1969). Seismicity (defined in Section 3), as compared with those based on
data following major events are treated with subjective the average occurrence rate (Naylor et al. 2009).
rules to characterize the clustered seismicity, including In general, all of the previously described declustering
spatial and temporal constraints, the number and size of algorithms either explicitly or implicitly use a measure of
triggered events, and constraints linked to the physical clustering (or dependence) between successive earth-
fault model under consideration. quakes. However, a comparison between these different
Most of the studies that assume a Poisson distribu- approaches is not readily possible because different hy-
tion of the earthquake catalog, particularly in probabi- potheses and models are used in each case. This limita-
listic seismic hazard assessment, use declustering tion motivated us to introduce a clustering measure that
algorithms to remove clustering from the original cata- rates each declustering algorithm according to its refer-
log. The residual catalog is then assumed to reflect ence background model, so that the performances of the
Poissonian earthquake occurrences. However, artifacts different algorithms can be evaluated. While a general
from the use of such residual catalogs have not been class of clustering measures, such as the Ripley K-
rigorously assessed (Beauval et al. 2006). Furthermore, function (Ripley 1976) and its variants, already exists,
the process of filtering the original catalog to obtain a their application is limited because estimation of the K-
Poissonian residual catalog is not unique, but a type of function is complicated by stationarity and edge effects
“Poissonianization” rather than “declustering” sensu (e.g., Cressie 1991). Hence, application of this type of
stricto (Luen and Stark 2012). clustering measure to earthquake data is presently lim-
One successful approach developed in the study of ited (e.g., Veen and Schoenberg 2006).
seismicity clustering is the use of stochastic declustering To further contribute to the debate about the merits
techniques, whereby events are triggered according to a of different declustering methods, this study focuses
J Seismol (2013) 17:1041–1061 1043

on scoring the methods using as a reference the paper was to measure directly the physical clustering
hypothesis adopted by each method rather than evalu- process. Hence, this part of the study is detailed and is
ating each method according to the subjective definition not concerned with artifacts arising from the use of
of what is an “aftershock,” “mainshock,” or “foreshock.” declustering algorithms as such algorithms are not
As such, we used time clustering at different ranges as the employed. Applications of our clustering measure are
measure of the relative distance between the empirical considered in Section 6, where the use of different
whole and background inter-event time distributions declustering algorithms is pivotal in comparing their
(Talbi and Yamazaki 2010). Relatively high values of performances.
our measure reflect the presence of clustering in the
corresponding inter-event time range, whereas low
values correspond to the dominance of background 2 Declustering
events. Our proposed new measure for recognizing
clustering offers a number of advantages over existing In this study, the terms “clustered,” “residual,” and
techniques, including the ability to theoretically measure “whole” seismicity correspond to “dependent,”
the clustering relative to a hypothetical background “background,” and “all” events in the catalog, respective-
model under general assumptions and applicability to ly. The term “residual” is preferred to “background”
non-Poissonian background seismicity. In application, when used with reference to the outputs of a declustering
each algorithm is rated by the measure according to the algorithm. In fact, the subjective algorithms used
deviation between its output and the hypothetical back- for declustering only provide a catalog that approxi-
ground model it assumes. As a result, the proposed mately (not rigorously) reflects background seismicity;
measure can be used to compare the performance of in particular, testing as to whether the declustered
different declustering algorithms, even if they use catalog is Poissonian should be carefully considered
different background assumptions. (Luen and Stark 2012). The term “background seismic-
This study builds on our former efforts to study ity” is the ideal, unknown, independent structure that we
inter-event time distributions (Talbi 2009; Talbi and seek to estimate in our clustering study. A preliminary
Yamazaki 2009, 2010). In particular, the mixed assessment of the background seismicity structure is
Weibull model of the inter-event time distribution obtained by processing with the algorithms of Gardner
(Talbi and Yamazaki 2010) is used in estimating the and Knopoff (1974), Reasenberg (1985), and Zhuang et
proposed clustering measure. al. (2002). These algorithms are based on space–
The remainder of this paper is organized as follows. time windows, physical fault models, and ETAS
Section 2 introduces the declustering approaches used models, respectively. This multifaceted approach is
to filter the catalogs from aftershocks, while Section 3 expected to minimize artifacts and focus on real
describes the mathematical formulations used to de- and dominant characteristics of earthquake clustering.
fine the clustering measure, in addition to discussing Additionally, the results may be used to test the perfor-
the shape and behavior of the proposed measure. mance of these declustering algorithms. Although over-
The sampling methodology and description of the views of seismicity declustering methods are described
earthquake catalogs are presented in Sections 4 and elsewhere (e.g., van Stiphout et al. 2012), this section
5, respectively, and Section 6 introduces the esti- provides a brief description of each of the declustering
mation of the proposed clustering measure using algorithms used later in this study.
real catalog data and analysis of clustering attenua- Gardner and Knopoff (1974) introduced a declustering
tion curves. Finally, Section 7 reports the main algorithm that uses the proximity of earthquakes in
results of this study. space and time as an indicator of clustering. For a
It is important to distinguish between the theoretical given catalog of data, events are ordered in decreasing
aspects and subsequent applications of our work in magnitude. Starting from the first event, space–time
this paper. The main part of this work introduces the windows are measured around each event in the catalog.
theoretical basis of our clustering measure and dis- The size L and duration T of each window vary with the
cusses its shape and behavior using two hypothetical magnitude M of the potential mainshock. The largest
models for background seismicity (Section 3 and the event in each window is identified as a mainshock, while
Appendix). The objective of this main part of the the others (foreshocks or aftershocks) are identified and
1044 J Seismol (2013) 17:1041–1061

removed. The window parameters are estimated using respectively. The earthquake databases used are
the following regressions: described in Section 5. Figures 1b and 2b show plots
of the residual seismicity obtained using the algorithm
logðT Þ ¼ a1 M þ b1 ð1Þ of Gardner and Knopoff (1974) for southern California
and Japan, respectively. Figure 1c presents seismicity
logðLÞ ¼ a2 M þ b2 ð2Þ data for southern California filtered using the algorithm
of Reasenberg (1985). In Fig. 1c, most of the clustering
For a specific earthquake catalog, the parameters structure is not evident in the filtered data, giving rise to
a1, b1, a2, and b2 can be estimated by interpolation of lower spatial density, which is concentrated close to
past aftershock zone extent and aftershock duration active faults. In contrast, there is less concentrated and
data. However, most studies use standard values of more diffuse southern Californian seismicity in Fig. 1b.
these parameters estimated from, for example, the mag- For Japan, the residual seismicity is mostly diffuse when
nitude–length–time data (M, L, and T, respectively) filtered using Gardner and Knopoff’s (1974) algorithm
given in Table 1 of Gardner and Knopoff (1974). (Fig. 2b). The main difficulty in the application of
The algorithm of Reasenberg (1985) assumes an in- Reasenberg’s (1985) algorithm is adjustment of the
teraction zone centered on each earthquake. Earthquakes algorithm parameters according to the seismicity of the
occurring within the interaction zone of a prior region. Typically, a selection of parameters is made after
earthquake are considered aftershocks. The zone is attempting different runs with varying parameters, and
dynamically modeled with spatial (Rfact) and temporal the various residual catalogs that are obtained from these
(τmax) parameters. The length scale Rfact is proportional runs are examined. Even after undertaking such a
to the source dimension, and the temporal scale τmax is procedure, clustering cannot be adequately removed and
determined using a heterogeneous Poisson process for background seismicity is usually overestimated (Hainzl et
aftershocks with rate l(t). Given t>0, the probability of al. 2006; Tibi et al. 2011). For this reason, we restrict our
observing n earthquakes in the time interval [t, t+C[ is testing of Reasenberg’s (1985) algorithm to the southern
given by California data where the original parameters have been
set, and for which a set of suitably chosen parameters has
elðtÞt ½lðtÞt n already been identified (Hutton et al. 2010).
Pð@ð½t; t þ t ½Þ ¼ nÞ ¼ ð3aÞ
n! The stochastic declustering procedure of Zhuang et
al. (2004) is based on space–time ETAS models
with l(t) following the Omori law (Ogata 1988) that use the following form of the con-
lðtÞ ¼ k ðt þ cÞp ð3bÞ ditional intensity:

ℵ is the process that counts the number of after- lðt; x; y; M jHt Þ ¼ lðt; x; yjHt ÞJ ð M Þ ð5aÞ
X  
shocks occurring in the time interval [t, t+τ[, whereas lðt; x; yjHt Þ ¼ μðx; yÞ þ k ðMi Þg ðt  ti Þf x  xi; y  yi jMi
k, c, and p are positive constants representing the i:ti <t
ð5bÞ
Omori law parameters. The waiting time interval, τw,
required to observe the next event with probability, P, l is the conditional intensity on the history of
in a given sequence of aftershocks is observation ℋt until time t, whereas μ, κ(M), g, f,
and J are the background intensity, the expected num-
t logð1  PÞ ber of triggered events from a magnitude M event, the
tw ¼ ð4Þ
102ðMmax mc 1Þ=3 probability distribution function of the occurrence
times of the triggered events, the spatial distribution
Mmax and mc in this equation are the largest mag- of the triggered events, and the magnitude probability
nitude in the sequence and completeness magnitude, distribution function, respectively. An important
respectively. The waiting time is constrained as falling ingredient in Eq. 5b is the modified Omori law, g
between the minimum and the maximum look-ahead (Utsu 1969; Yamanaka and Shimazaki 1990).
times (i.e., τmin ≤τw ≤τmax).
Figures 1a and 2a show the original seismicity data p1 t p
gðtÞ ¼ 1þ ð5cÞ
before declustering for southern California and Japan, c c
J Seismol (2013) 17:1041–1061 1045

Fig. 1 Epicenter distribution of whole (a) and residual seismic- structure is eliminated giving rise to lower spatial density, which
ity (b, c) in southern California for magnitude M≥2.5 events is concentrated close to active fault segments (b) and
within the period 1932–2010. The residual catalog is obtained Reasenberg’s (1985) algorithm with τmin =5, τmax =150, P=
using the algorithm of Gardner and Knopoff (1974), where a 0.90, and R fact =6, where a total of 26,417 events are identified
total of 16,848 events are identified as non-clustered out of as non-clustered and a similar picture of background seismicity
60,092 events in the whole catalog and most of the clustering is depicted (c)

     
In the above equation, p and c are the Omori law z i tj ; xj ; yj ¼ kðMi Þg tj  ti f xj  xi ; yj  yi jMi
parameters defined in Eq. 3b. ð7Þ
The principle behind this algorithm is the classifi-
cation of each event j as a triggered or background 8j ¼ 1  ρj ð8Þ
event using the following standardized intensity rates,
ρj and φj, respectively, obtained from Eq. 5b: This stochastic declustering algorithm simulates
X background and triggered events from the whole catalog
ρj ¼ ρi j ð6aÞ
i<j by scanning all events and treating them according to
8 the probabilities calculated using Eqs. 6–8. Hence, the
< z ðtj ;xj ;yj Þ  resulting catalog is not unique because it uses the like-
;j > i
ρij ¼ l tj ;xj ;yj jHt j ð6bÞ lihood of such probabilities. Figure 2c, d shows the
:
0 otherwise seismicity in Japan before and after applying the
1046 J Seismol (2013) 17:1041–1061

declustering algorithm of Zhuang et al. (2004). Accordingly, f can be divided into two components
Background seismicity is concentrated along the Japan obtained by filtering
Trench, whereas it is relatively diffuse at distances
farther from the Japan Trench. In contrast with the f ðt Þ  ð1  C ðt ÞÞ f ðt Þ þ C ðt Þ f ðt Þ ð11Þ
algorithm of Gardner and Knopoff (1974) (i.e., the
seismicity inside the dashed rectangle in Fig. 2a, b), Furthermore, dividing Eq. 10a by f(τ) allows one to
the overall structure of the spatial distribution of reduce the scaling effect of the distributions. For ex-
seismicity is preserved after the application of the ample, if e
f and e f B are the inter-event time probability
declustering method of Zhuang et al. (2004). density functions scaled using the inter-event time
mean t of the whole series, then
Z þ1
3 Mathematical background t¼ t f ðt Þ d t ð12aÞ
0
Assume seismicity to occur as a point process in space
and time (e.g., Cox and Isham 1980; Daley and 1e
f ðt Þ ¼ f ð t =t Þ ð12bÞ
Vere-Jones 1988). Within a finite region in space, the t
process is inferred to follow an inter-event time prob-
ability density function f, which is a perturbed form of 1e
f B ðt Þ ¼ f ð t =t Þ ð12cÞ
the probability density function fB of the background t B
seismicity component. It is then possible to write the
following equations: e is defined
The corresponding induced measure C
for scaled inter-event times as
f ðt Þ ¼ f B ðt Þ þ ½ f ðt Þ  f B ðt Þ ð9aÞ
e e
e ð t =t Þ  f ð t =t Þ  f B ð t =t Þ
C ð13Þ
f ðt Þ  f B ðt Þ þ " ðt Þ ð9bÞ e
f ð t =t Þ

The variable τ≥0 is the inter-event time and ε(τ)≈ It then follows that the two measures are equivalent
f(τ)−fB(τ) is the corresponding deviation from the in the sense that
background distribution. The proprieties of ε are
discussed further in part A of the Appendix. e ð t =t Þ ¼ C ð t Þ
C ð14Þ
It is important to note that ε is zero in the case of a
pure background process. For inter-event times τ≥ In Eqs. 12b and 13, the mixed inter-event time
0 with f(τ)>0, the deviation ε(τ) defines a standardized distribution, which has been defined for heteroge-
probabilistic measure C(τ) of the distance between the neous seismicity (e.g., Bak et al. 2002; Talbi and
whole and background inter-event time distributions Yamazaki 2010), is used as the scaled inter-event time
as follows: probability density function e f . This function has ap-
proximately the same shape for different time periods,
" ðt Þ f ðt Þ  f B ðt Þ
C ðt Þ   ð10aÞ magnitudes, and space scales (Talbi and Yamazaki
f ðt Þ f ðt Þ 2010). It should be noted that the local inter-event
time distribution fitted using a gamma distribution
In this case, both fB and ε appear as filtered versions (e.g., Fig. 2 of Corral 2003) is different and holds only
of the whole seismicity probability density function f, for stationary periods. It is important to stress this dif-
in which C and 1−C represent the role of time-varying ference because both are confusingly called the “scaling
multiplicative filters. law for earthquake inter-event time distribution” by
different authors. In our case, the probability density
" ðt Þ  C ðt Þf ðt Þ ð10bÞ
function ef can be analytically approximated using
previously published fitting models (Saichev and
f B ðt Þ  ð1  C ðt ÞÞ f ðt Þ ð10cÞ Sornette 2007; Talbi and Yamazaki 2010). As a result of
J Seismol (2013) 17:1041–1061 1047

Fig. 2 Seismicity map for Japan for magnitude M≥3.5 events catalog data used by Zhuang et al. (2004) are extended to 2009
registered during the period 1923–2010. a Original data includ- and used in schemes J4a-c (see the following c and d). c, d
ing 87,037 events. b Declustered catalog using the algorithm of Japan seismicity with magnitudes M≥4.2 registered within the
Gardner and Knopoff (1974) including 22,317 events. Dashed period 1923–2009 using the original data including 31,644
rectangle shows the area used by Zhuang et al. (2004) to test events (c) and the declustered catalog of Zhuang et al. (2004)
their stochastic declustering algorithm. In our study, the JMA including 12,753 events (d)

this, C depends mainly on the mean scaled background background model. In the following analysis, the approx-
distribution e
f B and can be calculated using a hypothetical imation of C used in application is given by Eq. 33 in the
1048 J Seismol (2013) 17:1041–1061

case of absolutely continuous probability measures P and event times close to τ0, within the range Iτ0 defined
P(⋅|B) and also in term of frequencies in Eq. 25. These by t 0  Δt  t < t 0 þ Δt. This probability is sub-
definitions are equivalent, and it is according to the sequently referred to here as the local background
properties of the measures P and P(⋅|B) whether to use frequency (LBF). LBF is estimated from the number
the exact expression (Eqs. 30b and 25) or the approxi- of background inter-event times τ within the range Iτ0
mation (Eqs. 33 and 25). It is important to note that divided by the total number of inter-event times in the
Eqs. 25, 30b, 33, and 38 are general and holds for a same range. Given that ν is the mean background
user-defined background model such as the gamma or frequency on the whole inter-event time scale, the
the inverse Gaussian distributions (Talbi and Yamazaki measure C(τ0) compares LBF and ν locally around
2010; Matthews et al. 2002). τ0. For a given ν, 1−C(τ0) is proportional to LBF, so
To provide a working definition of C, it is that low C(τ0) values correspond to high LBF and vice
necessary to define the probability P(B)=P(τ∈B) versa. C(τ0) is zero when the local background fre-
that an inter-event time τ belongs to the background. quency is equal to the mean ν, whereas it is positive
For a sufficiently large number of events, this probabil- when LBF is below the mean and negative when it
ity is comparable to the proportion ν of background exceeds the mean. Conversely, C(τ0) tends toward
events in the whole catalog, which is commonly called unity for very low background frequencies at the dis-
the background fraction; that is tribution tails. C(τ0) cannot be defined if no inter-event
time is registered in the vicinity of τ0 (i.e., when
P ðBÞ  v ð15Þ
Pðt 0  Δt  t  t 0 þ Δt Þ ¼ 0Þ.
In the following discussion, C has been calculated
For regions where aftershock activity is not domi-
analytically and plotted for a set of inter-event time models
nant, a simple biased nonparametric estimate of ν is
to assess its behavior. For this purpose, the use of the exact
the ratio of the first- to second-order moments of the
expression of C obtained from Eq. 30a is preferred over
inter-event times, i.e., inter-event time mean over var-
the use of the approximation given by Eq. 33 (part B of
iance (Molchan 2005; Hainzl et al. 2006).
the Appendix). As previously mentioned, the analytical
It is possible to distinguish two cases, where the
expressions of the cumulative distribution function for all
first is the general case when the cluster (non-back-
the inter-event times and its corresponding probability
ground) inter-event time distribution is unknown and
distribution function, which are required in Eqs. 30a and
the second is a specific case where it is driven from a
33 to calculate Pðt 0  Δt  t < t 0 þ Δt Þand f ðt0 Þ,
specific probability density function fN. In both cases,
respectively, can be found in, for example, Saichev and
the measure C is calculated explicitly.
Sornette (2007) and Talbi and Yamazaki (2010). In this
study, Poissonian and non-Poissonian models are tested
3.1 Case 1: Cluster inter-event time distribution
as background.
is unknown
Figure 3a shows a plot of two analytical fits of the
inter-event time probability distribution function f against
In this case, C can be defined in the neighborhood of
three exponential distributions with different means μB,
an inter-event time τ0 as
symbolizing the background distribution fB. For f, we
PðBjt 0  Δt  t < t 0 þ Δt Þ consider the mixed Weibull distribution MixWeib(c, p,
C ðt 0 Þ ¼ 1  ð16Þ
n k, θ) with c=0.2, p=1.5, k=0.5, and θ=0.7 (Eqs. 14, 28,
54, and 55 of Talbi and Yamazaki 2010).
The variables Δτ, ν, and PðBjt 0  Δt  t < t 0 þ " #
Δt:Þ are a positive time increment, the background f ðxÞ ¼ μ 
1 k
 Q1 ðxÞeðx=θÞ þ
1
Q2 ðxÞ
1
fraction, and the posteriori probability of background θΓ 1þ k
1 μ D ðx þ cÞ1þa
inter-event times, respectively. The derivation of   ð17aÞ
Eq. 16 is detailed in part B of the Appendix. ca c
Q1 ðxÞ ¼ 1 ðx þ cÞ1a þ kθk xk1
The posteriori conditional probability in the former ð1  aÞμD ð1  aÞμD
equation describes approximately the proportion of 2ca 1
þ ð17bÞ
local background inter-event times from all inter- μD ðx þ cÞa
J Seismol (2013) 17:1041–1061 1049

!
1 3.1.1 Example 1: Poisson background
a   Jk;θ ðxÞ
Q2 ðxÞ ¼ a c 1 ð17cÞ
θ Γ 1 þ 1k
In this example,
 
Z x et 0 =μB eΔt =μB  eΔt =μB
k C ðt 0 Þ ¼ 1  ð18Þ
Jk;θ ðxÞ ¼ eðt=θÞ d t ð17dÞ F ðt 0 þ Δ t Þ  F ðt 0  Δ t Þ
0

The derivation of Eq. 18 is detailed in part C of the


aca c1þa c1a Appendix. Although the analytical form of F can be
μD ¼ ðt max þ cÞ1a þ a 
1a ðt max þ cÞ a ð1  a Þ calculated in this case (Talbi and Yamazaki 2010), its
ð17eÞ substitution into Eq. 18 is not carried out here for the
sake of simplicity.
The variables μ and τmax are equal to 1 and 103, respec- The calculated C measures are plotted in Fig. 4a for
tively, while α is linked to p as α=1−1/p, 0<α<1. different μB values. For the mixed Weibull model, C
Alternatively, we use a rational interpolation to decreases gradually to reach its minimum for inter-
mimic the behavior of inter-event times at short and event times close to the background mean μB. In
1θ
long ranges, hðt Þ ¼ 1þθnt1þt 2þc , with n=0.9, θ=0.03,
contrast, for the interpolating function h, C decreases
and χ=0.7 (Eq. 76 of Saichev and Sornette 2007). irregularly to a minimum value that deviates from the
Figure 3b shows the plots of the probability distribu- background mean. At short inter-event times, C has
tion function f against a set of Weibull distributions values around unity, which correspond to very low
with different parameters. In most cases, departure proportions of background inter-event times compared
between the whole and the background probability with the overall series. At long inter-event times, C
distribution functions, f and fB, respectively, is increases rapidly to unity. This behavior is clearly
observed at the distribution tails, while the two proba- linked to the steep decrease in the exponential proba-
bility distribution functions are relatively similar around bility distribution function observed in Fig. 3a and is
the mean t. The deviation between f and fB observed in an unsurprising result given that the models for f
Fig. 3a, b is studied using the measure C in the following mimic real data with a non-exponential Weibull or
two examples. power law-like long tail (Talbi and Yamazaki 2010).

Fig. 3 Plots of the probability distribution function f of earth- magenta. The distribution f is coupled with plots of the exponential
quake inter-event times using a mixed Weibull distribution distribution with different means μB =2, 1, and 0.5, symbolizing
MixWeib(c, p, k, θ) with c=0.2, p=1.5, k=0.5, and θ=0.7 (Eq. Poissonian background occurrences (shown in red) (a), and plots
15 of Talbi and Yamazaki 2010), shown as a black solid curve, and of the Weibull distribution with different scale and shape param-
the rational interpolation hðt Þ ¼ 1þθnt
1θ
eters k=2 and 0.5, θ=2, 1, and 0.5, symbolizing Weibull back-
1þt 2 þc with n=0.9, θ=0.03,
and χ=0.7 (Eq. 76 of Saichev and Sornette 2007), shown in ground occurrences (shown in blue and red) (b)
1050 J Seismol (2013) 17:1041–1061

Fig. 4 Theoretical plots of the clustering measure C calculated mean μB =0.5, 1, and 2, respectively. b, c The curves correspond
using the mixed Weibull distribution (in black) and the rational to a Weibull background distribution with scale parameter θ and
interpolation h (in magenta). a Dotted, solid, and dashed curves shape parameter k=2 (b) and k=0.5 (c). Dotted, solid, and
correspond to an exponential background distribution with dashed curves correspond to θ=0.5, 1, and 2, respectively

Figure 3a clearly supports the presence of long-range in Talbi and Yamazaki (2010), it is not reproduced here
clustering in real data. The two maxima in C visible for for the sake of simplicity.
the interpolating function h with μB =1 (solid magenta The calculated C measures are plotted in Fig. 4b, c
curve) are caused by the failure of h to account for the for different values of the parameters k and θ. In the case
behavior of inter-event times near the mean (Talbi 2009). k=2, which corresponds to the Rayleigh distribution, C
is plotted in Fig. 4b. Its behavior is quite similar to
3.1.2 Example 2: Weibull background Fig. 4a, but with a sharper and deeper V-shaped behav-
ior. The minimum C value for all curves is quite close to
In this example, the scale parameter θ. Given that θ and μB are compa-
k k rable (θ=0.5, 1, and 2 correspond to μB =0.44, 0.89, and
e½ðt 0 Δt Þ=θ  e½ðt 0 þΔt Þ=θ 1.77, respectively), the minima are also close to the
C ðt 0 Þ ¼ 1  ð19Þ
F ðt 0 þ Δ t Þ  F ðt 0  Δ t Þ mean background inter-event time μB. The case
for k=0.5 is shown in Fig. 4c. For τ≤1, C is quite
The derivation of Eq. 19 is detailed in part D of the stable, with values close to zero, particularly for the
Appendix. Although the analytical form of F is provided mixed Weibull distribution (black curves). This type of
J Seismol (2013) 17:1041–1061 1051

behavior reflects the dominance of the background pro- ν can be estimated from the normalization condition
cess at short ranges, with a relatively constant LBF close imposed by weights in Eq. 20 as follows:
to the mean ν within each inter-event time bin. At longer
μ μ
inter-event times (i.e., τ≥1), it is possible to distinguish ð1  n Þ þn ¼1 ð21Þ
μN μB
two cases corresponding to θ<1 and θ≥1. In the first
case, where θ=0.5, which corresponds to μB =1, C is
quite stable, varying around zero for a wide range of Subsequently, for μ>0 and μB −μN >0,
inter-event times (dotted curves). This stability reflects  
the contiguity of the probability distribution functions μB μ  μN
n¼ ð22Þ
fB and f and the fact that fB (the stretched exponen- μ μB  μN
tial distribution) is successful in capturing the long-
term behavior of the whole inter-event time series. In Combining Eqs. 10a, 20, and 22, an expression for
the second case, where θ≥1 (i.e., θ=1 and 2, which the approximation of C can be derived as
corresponds to μB =2 and 4, respectively), C de-
1
creases to a minimum in the case of the linear C ðt Þ  1      ð23Þ
μB μ fD ðt Þ μμN
interpolation h. The minimum C value is much μB μN fB ðt Þ þ μB μN
higher than the mean μB value in this case. The
differences between the background and the whole
distributions at short and long time ranges indicate This expression can be simplified to
time clustering, whereas the minimum C value
1
attained at around the mean background inter-event C ðt Þ  1  ;
time μB identifies the maximum observed LBF, and ð1  bÞ ffNB ððttÞÞ þ b
ð24Þ
hence the dominance of inter-event times from back- μ  μN
ground seismicity. with b ¼
μB  μN
Finally, it is important to note the typical V-shaped
behavior observed in Fig. 4a, b. For very large μB
values that are much greater than the mean t, C Equation 24 shows that for a given β, C tracks the
reaches a minimum outside the observable range; con- same variations (increase and decrease with τ) as the
sequently, C appears to decrease along the whole inter- ratio of cluster to background probability distribution
event time range and the V-shaped behavior cannot be functions.
observed.

3.2 Case 2: Cluster inter-event time distribution 4 Inter-event time sampling and inference
is known
The approach adopted in this study uses earth-
In this case, the probability distribution function f can quake random sampling (ERS) to produce inter-
be separated into two components as follows: event time series. The algorithm performs an iter-
ative sampling by computing seeds from a network
μ μ
f ðt Þ ¼ ð1  n Þ fN ðt Þ þ n fB ðt Þ ð20Þ of local disks with radius R in each run (Talbi and
μN μB
Yamazaki 2010). R is increased with larger mag-
nitudes when sampling is judged to be poor.
The variables ν, μB, μN, and μ are the back- Subsequently, the obtained time series are mixed
ground fraction (Eq. 15), the mean inter-event time and scaled using the arithmetic mean inter-event
of the background, and the clustered and whole time t obtained from the whole series. In addition
seismicity components, respectively. fB and fN are to this approach being able to reveal an approxi-
the conditional probability distribution functions of mately unique shape of the inter-event time distri-
the background and clustered seismicity compo- bution (e.g., Bak et al. 2002; Christensen et al.
nents, respectively. The derivation of Eq. 20 is 2002; Corral 2007), such a scaling procedure provides a
detailed in part E of the Appendix. The proportion standard dimensionless reference where the terminology
1052 J Seismol (2013) 17:1041–1061

“short term” versus “long term” can be used relative to the catalog for the period 1932–2010 was compiled from
the mean value. “Short term” is used in reference to the Southern California Seismic Network (http://
inter-event times with t and “long term” is used in www.data.scec.org/ftp/catalogs/SCSN/). The region be-
reference to inter-event times with t >> t. The unique- tween latitudes 32°–37° N and longitudes 114°–122° W
ness of the scaled inter-event time distribution f for was considered (Fig. 1a). Seismicity in southern
different regions and completeness magnitudes was pre- California is dominated by mainshock–aftershock se-
viously used in the theoretical calculation of C in quences. Dense clusters are formed by aftershocks con-
Section 3. In this section, the same scaled distribution centrated on mainshock rupture zones, in addition to
is used for consistency. We also use the mean inter-event individual swarms. Some of the largest earthquake se-
time scaling because it is more stable than the mean quences recorded were the 1952 Mw =7.5 Kern County,
event rate scaling. It should be noted that the occurrence the 1992 Mw =7.3 Landers, and the 1999 Mw =7.1 Hector
of an extreme event adds a large number to the sum of Mine earthquakes, which all ruptured slow-slipping
past events rates and a small amount to the sum of past faults east of the San Andreas Fault. These events in-
inter-event times (Naylor et al. 2009). duced long-lasting aftershocks that continue today,
The degree or effect of clustering is calculated using which are distributed over tens of kilometers within the
the measure C as defined in Eq. 10a. As the independent corresponding mainshock epicentral zones (Hutton et al.
seismicity structure is unknown, fB cannot be estimated 2010). The catalog also includes smaller but still damag-
reliably. However, the measure C allows us to constrain, ing earthquakes such as the 1933 Mw =6.4 Long Beach,
at least in part, how time correlation between events is the 1971 Mw =6.7 San Fernando, and the 1994 Mw =6.7
linked. Even in the case where the declustering method Northridge earthquakes. Hutton et al. (2010) showed that
fails to remove all correlations, C should be able to the magnitude of completeness of the SCSN catalog
identify at least large and small deviations between the exceeds roughly 3.2 and 1.8, starting from 1932 and
whole and background distributions along different 1981, respectively, whereas it reaches 1.0 within the core
inter-event timescales. of the network coverage area.
In Section 6, C is computed using the inter-event In our case studies, the magnitudes of complete-
time frequencies fr and frB, corresponding to the em- ness, the corresponding time periods, together with the
pirical whole distribution obtained from the original number of events and depths are shown in Table 1 for
data and the residual empirical distribution obtained both southern California and Japan. The letters SC and
using a chosen declustering approach, respectively. J in the first column refer to southern California and
The frequencies fr and frB are calculated given a par- Japan, respectively. In the case of southern California,
ticular inter-event time bin I; then, for each inter-event the magnitude of completeness and their correspond-
time τ∈I, C(τ) is approximated as ing time periods (Table 1) are taken from Talbi and
Yamazaki (2009). These magnitudes of completeness
frB for southern California are higher than those given by
C ðt Þ  1  ; fr > 0 ð25Þ
fr Hutton et al. (2010).
For Japan, we used the Japan Meteorological
Since broad timescales are involved (from seconds Agency (JMA) catalog covering the period 1923–
and minutes to years), an appropriate bin size along the 2010 and within the region between latitudes 24°–
inter-event time axes can be obtained by considering the 50° N and longitudes 122°–152° E. This data
logarithmic scale (e.g., Bak et al. 2002) or an exponen- source is a modern seismic network operated and
tially increasing series ci, where i is the label of consec- maintained by the JMA. Most seismicity is typi-
utive bins and c>1 (e.g., Corral 2004, 2005). In this cally concentrated along the Japan Trench in the
study, we set c=1.5; other choices give similar results. subduction zone between the Pacific and Eurasian
plates (Fig. 2a). A recent study by Nanjo et al.
(2010) showed that relatively low completeness
5 Earthquake data magnitudes, mc ≥1.9, characterize the mainland re-
gion, which is consistent with the high density of
Earthquake catalog data from southern California and seismic stations in this region; in contrast, com-
Japan were used in this analysis. For southern California, pleteness magnitudes are high in offshore areas
J Seismol (2013) 17:1041–1061 1053

Table 1 Parameters of the sampling schemes

Scheme Period mc Depth N R (km) Figure Dec

SC1 1990–2010 2.5 All 27,693 50 Fig. 5b, d GK, R


SC2 1947–2010 3.5 All 6,308 50 Fig. 5b, d GK, R
SC3a 1932–2010 4.5 All 781 50 Fig. 5b GK
SC3b 1932–2010 4.5 All 781 100 Fig. 5d R
J1a 1990–2010 3.5 All 47,686 50 Fig. 5a GK
J1b 1990–2010 3.5 <100 km 39,718 50 Fig. 5a GK
J2a 1975–2010 4.5 All 12,741 50 Fig. 5a GK
J2b 1975–2010 4.5 <100 km 10,985 50 Fig. 5a GK
J3a 1923–2010 5.5 All 3,864 100 Fig. 5a GK
J3b 1923–2010 5.5 <100 km 3,180 100 Fig. 5a GK
J4 1926–2009 4.2 <100 km 31,644 50 Fig. 5e Z

Period, mc, N, R, and Dec denote the start and end year of events considered in the scheme, the magnitude of completeness, the number
of all events with magnitudes M≥mc, the ERS sampling radius, and the declustering method used, respectively. These parameters are
grouped and labeled according to the sampling schemes defined in the first column. SC and J correspond to southern California and
Japan, respectively. GK, R, and Z correspond to the declustering algorithms proposed by Gardner and Knopoff (1974), Reasenberg
(1985), and Zhuang et al. (2004), respectively. The last two columns (Figure and Dec) list the figures in this paper that correspond to the
use of each of the sampling schemes and declustering methods. For example, for SC1, Fig. 5b corresponds to the GK algorithm,
whereas Fig. 5d corresponds to the R algorithm

where few seismic stations are present. As is the background processes, and in identifying differences
case for southern California, the magnitudes of between declustering algorithms, are discussed. For this
completeness listed in Table 1 for Japan, excluding discussion, different declustering algorithms are applied
scheme J4, were derived in our former study to reduce possible artifacts and then C is computed in
encompassing time periods up to 2005 (Talbi and each case, combined with its corresponding sampling
Yamazaki 2009). Here, the time periods in Table 1 scheme as given in Table 1. Only major and persistent
of Talbi and Yamazaki (2009) were extended to trends are used for the characterization of the clustering
2010. In addition to the study of Talbi and and background seismicity structures. Although the in-
Yamazaki (2009), the schemes J1a–b, J2a–b, J3a– terpretation of the results is sensitive to artifacts, it still
b have magnitudes of completeness and time pe- clearly distinguishes between the clustering behaviors
riods comparable with the magnitudes of complete- produced by the different declustering techniques.
ness derived by Nanjo et al. (2010). Scheme J4 Given that C typically decreases with increasing
corresponds to the JMA catalog data used in the inter-event times, it is convenient to introduce the term
study of Zhuang et al. (2004), but with the time “clustering attenuation curves” (CACs) in reference to
period extended in our study to 2009. Scheme J4 the behavior of C. Figure 5a, b shows the CACs obtained
covers the region with latitudes 30°–46° N and for Japan and Southern California seismicity, respective-
longitudes 128°–148° E. This scheme is only used ly, using the algorithm of Gardner and Knopoff (1974).
in this study to test the stochastic declustering The last column of Table 1 identifies the schemes and
algorithm. parameters used in each figure. In Fig. 5a, the CACs for
shallow depth events (d<100 km) are plotted with red
markers, and the theoretical fit of C is shown using the
6 Results and analysis mixed Weibull distribution as the whole distribution f and
an exponential background distribution fB with mean
The measure C is computed and plotted for each sam- μB =4. This fit suggests a background fraction of about
pling scheme listed in Table 1, and then the usefulness 25 %, which is comparable with the ratio of 22,317
of the results in the characterization of clustering and independent events out of 87,037 total events identified
1054 J Seismol (2013) 17:1041–1061

by the declustering algorithm. An increase in C is not fluctuations at long range and cannot be adequately
observed in the real data because of the shortage of data fitted as in Fig. 5a. However, it is possible to consider
at long range, ðt > 10t Þ. The data in Fig. 5b show high the best fit to the data with the models presented in
J Seismol (2013) 17:1041–1061 1055

R Fig. 5 Plots of the clustering measure C calculated using the


test the fitting models in Fig. 5b, we use the mean
algorithm of Gardner and Knopoff (1974) for Japan (a) and
southern California (b). Red markers in (a) show C calculated square residual function defined as
for shallow (d<100 km) events. Dashed and dotted curves in (b)
show theoretical fits calculated using the mixed Weibull distri- 1 X  2
bution as the whole inter-event time distribution and an expo- RMSðt Þ ¼ b
i=xi t C ðxi Þ  C ðxi Þ ð26Þ
nm
nential background distribution with μB =2 and 3, respectively.
Solid curves in (a) and (b) show the fits calculated using the
mixed Weibull distribution as the whole inter-event time distri- C and Cb are the measure predicted by the models in
bution and an exponential background distribution with μB =4 Fig. 5b and the measure calculated from Eq. 25 for the
and a mixed exponential−Weibull distribution, respectively. c
Plot of the residual mean squares, RMS, corresponding to the fit schemes in Fig. 5b, respectively. The variables n and
in (b). Blue, red, and black markers correspond to the residuals m are the sample size and the number of free param-
of the dashed, dotted, and solid curves in (b). d Plots of the eters in the model, respectively. In our case, we sup-
clustering measure C calculated using the algorithm of pose f fixed and count the number of free parameters
Reasenberg (1985) for southern California using the parameters
τmin =5, τmax =150, P=0.9, and Rfact =6. Solid line shows loga- in the model for fB. Therefore, m=1 when fB is expo-
rithmic fits to the data between 102 t and t , whereas the 95 % nential and m=4 when fB follows the mixed exponen-
limiting curves are shown by dashed lines. e Plots of the tial–Weibull model. From the RMS plot in Fig. 5c, we
clustering measure C calculated using the algorithm of Zhuang confirm that the best fit is obtained using an exponen-
et al. (2004) for Japan seismicity with magnitudes M≥4.2 within
the period 1923–2009. The solid curve shows the mean C curve tial background distribution with μB =3 (red markers)
obtained from 100 runs, whereas the dashed curves show the for t  t, while for t > t, the use of the mixed expo-
95 % confidence limits of the mean curve nential–Weibull background distribution (black
markers) provides the best fit to our data. In Fig. 5b,
Section 3. Data in Fig. 5b are fitted in a similar the slight increase of C visible at long range for
fashion to Fig. 5a, but with μB =2 and 3 as dashed scheme SC2 supports the validity of the theoretical
and dotted curves, respectively. This model suggests fit. It should be noted that, in general, C trends to zero
a background fraction ranging between 33 and 50 %, around the mean inter-event time t in the case of
although only about 25 % of the data are identified as shallow depth events (red markers in Fig. 5a), and
background by the declustering algorithm (i.e., otherwise at shorter times (black markers in Fig. 5a,
16,848 events out of 60,092 total events). This in- b). Exceedingly small values of C indicate that LBF is
consistency between the real data and the fitted mod- reaching the mean ν (LBF≈ν). For t  0:1t, C(τ) has
el for t > t can be explained by the deviation of the values around 1, indicating a very low background
residual events identified by the declustering algo- frequency. The clustering component is expected to
rithm from the assumed Poissonian behavior. Let us be dominant because of the very low background
show that the behavior of C for t > t can be frequency registered at a short inter-event time range.
reproduced using a mixed exponential–Weibull back- For t  t in Fig. 5a and t  0:1t in Fig. 5b, C(τ)
ground distribution, i.e., fB~ pr ExpðμÞ þ ð1  pr Þ decreases rapidly to values less than −1, which
Weibðk; θÞ; 0  pr  1. The solid line in Fig. 5b corresponds to a background inter-event time fre-
corresponds to a mixed exponential–Weibull back- quency more than twice the whole inter-event time
ground distribution with pr=0.45. It fits quite well frequency. In this case, the clustering is expected to
the data for t > t. The whole data set is the best fitted be low and to decrease with increasing LBF. Both
by an exponential background with μB = 3 (dotted positive and negative C values indicate the pres-
curve) for t  t and the mixed exponential–Weibull ence of clustered pairs of events in the correspond-
background distribution (solid curve) for t > t. A ing time ranges.
consistent way to test point process fitting is to use In order to test the consistency of the former obser-
diagnostic methods applicable to point processes vations, C has been calculated using the algorithms of
with known conditional intensity (Stoyan and Reasenberg (1985) and Zhuang et al. (2004) for south-
Grabarnik 1991; Lawson 1993; Baddeley et al. ern California and Japan, respectively. Figure 5d is a
2005, 2008, 2011; Zhuang 2006). However, in our graph of the CACs for southern California calculated
case, these methods are not applicable because the using Reasenberg’s (1985) declustering algorithm
conditional intensity is not defined. Alternatively, to with the parameters τmin =5, τmax =150, P=0.90, and
1056 J Seismol (2013) 17:1041–1061

Rfact =6. These values are slightly below the middle Poissonian residual inter-event time distribution. The
range of acceptable values used by Hutton et al. correlation identified using our clustering measure
(2010) for the SCSN catalog. This choice of parame- varies with the declustering algorithm used. While the
ters accounts for low-magnitude events since the algorithm of Gardner and Knopoff (1974) shows a
declustering algorithm is applied to M≥2.45 events strong correlation with a weak background component
instead of M≥2.95 events, as used by Hutton et al. at short ranges (Fig. 5a, b), time correlation decreases
(2010). The same periods and magnitude of complete- more rapidly at short term and saturates at long ranges
ness used in Fig. 5b are considered with a sampling with an almost constant LBF when the algorithms of
radius R=100 km instead of R=50 km for M≥4.5 Reasenberg (1985) and Zhuang et al. (2004) are used
events. A linear fit with 95 % confidence limits shows (Fig. 5d, e). This discrepancy can be explained by the
that C decreases logarithmically for 102 t  t  t. basic assumptions used by each of the algorithms.
After the mean, C appears to be relatively constant, with Reasenberg (1985) and Zhuang et al. (2004) assume
values between approximately −0.8 and −0.6. This be- an Omori law non-homogenous Poisson process for
havior is partly a function of the use of Reasenberg’s aftershocks (Eqs. 3b and 5c), whereas Gardner and
(1985) algorithm for aftershocks (assuming a non- Knopoff (1974) do not use such assumption. The algo-
homogeneous Poissonian with an Omori law rate, Eqs. rithm of Gardner and Knopoff is removing all events
3a and 3b), which results in a power law inter-event time that occur close in time around each event (i.e., digging
distribution (Utsu et al. 1995; Shcherbakov et al. 2005, time holes), resulting in the overestimation of clustering
2006; Yakovlev et al. 2005). Figure 5d shows the mean (aftershocks), especially at short time ranges.
CAC obtained using 100 stochastically declustered cat- Finally, it is interesting to note the typical behavior of
alogs with corresponding 95 % confidence limit curves. C in both Fig. 5d, e. At short range, Omori law holds for
This figure is highlighted for two reasons. Firstly, the aftershocks worldwide and, thus, constitutes a bench-
CAC demonstrates the clustering in the ETAS model, mark hypothesis. The Omori law aftershock model is
which is one of the most successful approaches previ- more than a simple hypothesis; it is an empirical law
ously used to model seismicity. Secondly, the stochastic fitting broad aftershock sequences worldwide. If we
declustering method does not use any subjective param- believe the Omori law, C should typically decrease reg-
eters, but objectively estimates them from the observed ularly (almost linearly) on the semi-logarithmic scale
seismicity. In Fig. 5e, the plot of C shows three different governed by the aftershocks that constitute most of the
types of behavior. A rapid logarithmic decrease in C that clustered events. At long range, C should be more stable
is similar to that observed in Fig. 5d is evident for or vary slowly. Figure 5a, b relative to the algorithm of
t  0:1t. C then shows a slower decrease for Gardner and Knopoff are more fitting our model with
0:1 t  t  t. At long ranges, a stationary-like regime Poissonian background, but are missing important infor-
appears to exist, but it is masked by marked fluctuations mation about the clustered events that is the Omori law.
in C. As is the case in Fig. 5d, the behavior of C for inter- In consequence, the algorithm of Gardner and Knopoff is
event times t  0:1t is linked to the Omori law hypoth- deficiently overestimating the clustering at short range
esis for cluster events in the ETAS. This illustrates the and underestimating the clustering at long range. The
regular disruption of the clustering structure and the fact that the different behaviors of our measure are linked
gradual settling of a relative increase in background to the assumptions made for the aftershock models in
inter-event time frequency. LBF reaches the mean ν at each algorithm is not a limitation since our measure can
around t  0:1t as C trends to zero. deal with any available hypothetical aftershock model.
Finally, the CACs plotted in Fig. 5a, b can be
modeled quite successfully using the mixed Weibull
model with exponential background distribution, with 7 Concluding remarks
the exception of the discrepancy observed in Fig. 5b at
long ranges, which is fitted using a mixed exponential– The first objective of this study was to introduce a
Weibull background distribution. The need of such clustering measure for seismicity by assuming a given
mixed background distribution can be explained by the background seismicity model. The clustering measure
failure of the declustering algorithm to closely model is simply defined as the standardized difference
J Seismol (2013) 17:1041–1061 1057

between the whole and background inter-event time are grateful to the Japanese Meteorological Agency and the South-
ern California Seismic Network for providing the catalog data used
frequencies. The proposed clustering measure does
in this study. The authors thank Rodolfo Console from INGV, an
not make any preconceived assumptions about the anonymous reviewer and the editor Torsten Dahm for their
nature of the clustered seismicity, but instead describes comments, which improved an earlier version of the manuscript.
it using a simple clustering attenuation curve. In par-
ticular, the Omori law assumed by the methods of
Reasenberg (1985) and Zhuang et al. (2004) induces Appendix
a linear decrease in the CACs (Fig. 5d, e). The pro-
posed clustering measure is then used in the analysis A. Proprieties of the deviation ε
of earthquake clustering at different inter-event time-
scales and for the interpolation of the results to describe Earthquake clustering is typically observed after the
clustered and background seismicity. Our results show occurrence of intermediate- to high-magnitude events.
that the clustering measure successfully quantifies local A series of events called “aftershocks” that occur closely
perturbations of seismicity produced by time clustering. in space and time are triggering phenomena. Our general
In particular, it distinguishes between a strong clustering hypothesis assumes that the overall pattern of seismicity
at short inter-event time ranges versus a weak and less can be described by a stable seismicity component
apparent clustering at long time ranges that is dependent overprinted by local perturbations that reflect clustering
on the declustering algorithm used. in space and time. The stable component is linked to
The second objective was to use our clustering some background process reflecting regional constraints
measure to compare the results obtained with three that are mainly tectonically driven seismicity, whereas
different declustering methods. The algorithms of the local perturbations produce clustering through the
Reasenberg (1985) and Zhuang et al. (2004), which time-dependent relaxation of the Earth’s crust.
assume an Omori-type aftershock decay, show that The deviation ε has the noise propriety that sums to
time clustering decreases more rapidly as a power zero.
law decay at short inter-event times and saturates at Z þ1
long inter-event times. These different behaviors are E ðC ðt ÞÞ ¼ CðsÞf ðsÞ d s ¼ 0 ð27Þ
linked to the assumptions made for the aftershock 0
models in each algorithm. The algorithm of Gardner
and Knopoff (1974), which assumes a finite space– These perturbations contribute to the observation of
time range of aftershocks, shows a strong time clus- clustering at different time ranges, with the clustering
tering with weak background seismicity at short inter- effect or extent calculated using the measure C as
event times. If we believe the Omori-type decrease for defined in Eq. 10a.
the number of events after the mainshock, the algo- A simple case where the deviation ε can be calcu-
rithm of Gardner and Knopoff (1974) only removes lated is the case of a simple mixed distribution be-
events close in space and time, resulting in the tween background and cluster events. In such a case,
overestimation of clustering at short range and under- Eq. 20 holds and f can be written as
estimation of clustering at long range.
In addition to the characterization of earthquake f ðt Þ ¼ bfB ðt Þ þ ð1  b ÞfN ðt Þ ð28aÞ
time clustering over different inter-event time ranges,
our proposed clustering measure is particularly useful With α defined as in Eq. 24,
in the detection of long-term clustering. Future work
f ðt Þ ¼ fB ðt Þ þ ð1  bÞ½fN ðt Þ  fB ðt Þ ð28bÞ
using this approach could be extended to the study of
space clustering and the combination of space and time
clustering analysis to construct a stochastic declustering Subsequently, ε can be derived from Eqs. 9b and 28b.
strategy. "ðt Þ  ð1  b Þ½fN ðt Þ  fB ðt Þ ð29Þ

Acknowledgments This work was supported by a fellowship Equation 29 shows that ε is proportional to the differ-
from the Japanese Society for the Promotion of Science. The authors ence between the two component distributions fN and fB.
1058 J Seismol (2013) 17:1041–1061

B. Derivation of Eq. 16

In terms of probability, the measure C can be defined


in the vicinity of an inter-event time τ0 as follows:

1
C ðt 0 Þ ¼ ½Pðt 0  Δt  t < t 0 þ Δt Þ  Pðt 0  Δt  t < t 0 þ Δt j BÞ ð30aÞ
P ðt 0  Δ t  t < t 0 þ Δ t Þ

Where Δτ > 0 is a time increment for which In this case, the following approximation is
Pðt 0  Δt  t < t 0 þ Δt Þ > 0. Simplifying Eq. 30a obtained:
gives
f B ðt 0 Þ
Pðt 0  Δt  t < t 0 þ Δt j BÞ C ðt 0 Þ  1  ; f ðt 0 Þ > 0 ð33Þ
C ðt 0 Þ ¼ 1  ð30bÞ f ðt 0 Þ
P ðt 0  Δ t  t < t 0 þ Δ t Þ
Equation 30b gives the exact expression C(τ0) that Equation 33 is equivalent to Eq. 10a for τ=τ0.
is used in parts C and D of the Appendix. In practice, C(τ0) is estimated using the whole and
If the probability P and its conditional form Pðj BÞ background inter-event time frequencies in the vi-
are absolutely continuous and the increment Δt is cinity of τ0 instead of f(τ0) and fB(τ0). In fact, our
small enough, then two probability distribution func- clustering measure simply compares the back-
tions, f and fB, exist such that ground and whole frequencies along the inter-
event time axes.
Pðt 0  Δt  t < t 0 þ Δt Þ  f ðt 0 Þ Δt ð31Þ
Alternatively, a useful expression of C(τ0) can be
calculated in terms of the background fraction ν. The
Pðt 0  Δt  t < t 0 þ Δt j BÞ  fB ðt 0 Þ Δt ð32Þ conditional probability in Eq. 30b is equivalent to

PðBjt 0  Δt  t < t 0 þ Δt ÞPðt 0  Δt  t < t 0 þ Δt Þ


P ðt 0  Δ t  t < t 0 þ Δ t j B Þ ¼ ð34Þ
PðBÞ

Finally, the expression of the measure C in Eq. 16 is 1 t=μB


fB ðt Þ ¼ e ð35aÞ
obtained from Eqs. 15, 30b, and 34. μB
t
 0
FB ðt 0 Þ ¼ 1  e μB ð35bÞ
C. Calculation of C in the case of Poissonian
background seismicity (Eq. 18) The variable μB is the mean inter-event time for
background events.
This example is important because it should mimic C The probabilities in Eq. 30b are calculated
calculated using real residual catalogs, which are as- using the cumulative distribution functions F and
sumed to be Poissonian. In this case, the background FB of the whole and background inter-event times,
inter-event time distribution is exponential. f B ∼ respectively.
Exp(μB), with a probability distribution function fB P ðt 0  Δ t  t < t 0 þ Δ t j B Þ
and a cumulative distribution function FB, respectively,
written as ¼ FB ðt 0 þ Δt Þ  FB ðt 0  Δt Þ ð36Þ
J Seismol (2013) 17:1041–1061 1059

P ðt 0  Δt  t < t 0 þ Δt Þ Similarly to the former case, C can be calculated


using the whole inter-event cumulative distribution
¼ F ðt 0 þ Δt Þ  F ðt 0 þ Δt Þ ð37Þ function F as given in Eq. 19.
Substituting Eqs. 36 and 37 into Eq. 30b yields
E. Derivation of Eq. 20
FB ðt 0 þ Δt Þ  FB ðt 0 þ Δt Þ
C ðt 0 Þ ¼ 1  ð38Þ If we assume that seismicity can be separated into
F ðt 0 þ Δt Þ  F ðt 0 þ Δt Þ
background and clustered components both described
by corresponding independent processes, the probabil-
Finally, by using Eq. 35b, the expression of the ity P(τ1 >τ) that the time to the next event from a fixed
measure C is that given in Eq. 18. origin 0 exceeds a given time τ>0 can be calculated
using the total probability theorem. This is given by
D. Calculation of C in the case of Weibull background the sum of the probabilities for the two components.
seismicity (Eq. 19)
P ðt 1 > t Þ ¼ P ðt 1 > t j BÞ P ðBÞ
The Weibull distribution (Weibull 1951) is able to
þ P ðt 1 > t j N Þ P ðN Þ ð42Þ
capture the long-term behavior of inter-event
times. For example, it has been used in a number B and N correspond to the following events:
of recurrence time models for large earthquakes B: “A background event occurs.”
(e.g., Newman et al. 2005; Yakovlev et al. 2006; N: “A clustered event occurs.”
Turcotte et al. 2007; Zoller and Hainzl 2007). It is
used here as a model for the background inter- P ðt 1 > t j BÞ and P ðt 1 > t j N Þ are the conditional
event time distribution. In this case, the probability probabilities of the forward recurrence time τ1 on the
distribution function fB and cumulative distribution events B and N, respectively (i.e., the probabilities that
function FB are starting from an arbitrary time, there occur no back-
ground events or no cluster events, respectively, in the
fB ðt Þ ¼ kθk xk1 eðx=θÞ
k
ð39aÞ following time interval of length t).
In this case, P(B) is the former background fraction
defined in Eq. 15.
k
FB ðt 0 Þ ¼ 1  eðt 0 =θÞ ð39bÞ PðBÞ ¼ n; PðN Þ ¼ 1  n ð43Þ
Applying the linear derivation operator to both
The variables k and θ>0 are the shape and scale sides of Eq. 42 gives
parameters, respectively. The Weibull distribution is
d P ðt 1 > t Þ d Pðt 1 > t j BÞ
noted hereafter Weib(k,θ), and we write fB ∼Weib(k,θ) ¼n þ ð1  n Þ
dt dt
in reference to its density fB defined in Eq. 39a. The
mean μB of the Weibull distribution is a function of the d P ðt 1 > t j N Þ
 ð44Þ
Weibull parameters. dt

  The derivatives in Eq. 44 can be obtained using Palm–


1 Khintchine equations (Cox and Isham 1980; Daley and
μB ¼ θ Γ 1þ ð40Þ
k Vere-Jones 1988). This also holds for the background and
clustered and whole seismicity components.
Using Eq. 39b, the following conditional probabil- d Pðt 1 > t j BÞ 1  DB ðt Þ
¼ ð45aÞ
ity is calculated: dt μB

Pðt 0  Δt  t < t 0 þ Δt j BÞ
d P ðt 1 > t j N Þ 1  D N ðt Þ
k
¼ e½ðt 0 Δt Þ=θ  e½ðt 0 þΔt Þ=θ
k
ð41Þ ¼ ð45bÞ
dt μN
1060 J Seismol (2013) 17:1041–1061

d P ðt 1 > t Þ 1  Dðt Þ Cox DR, Isham V (1980) Point processes. Chapman and Hall,
¼ ð45cÞ London
dt μ Cressie N (1991) Statistics for spatial data. Wiley, New York
Daley DJ, Vere-Jones D (1988) An introduction to the theory of
The substitution of the derivatives defined in Eqs. 45a,
point processes. Springer, New York
45b, and 45c into Eq. 44 yields Davis SD, Frohlich C (1991) Single-link cluster analysis, syn-
thetic earthquake catalogs, and aftershock identification.
1  D ðt Þ 1  D N ðt Þ 1  D B ðt Þ
¼ ð1  n Þ þn ð46Þ Geophys J Int 104:289–306
μ μN μB Frohlich C, Davis SD (1990) Single-link cluster analysis as a
method to evaluate spatial and temporal properties of earth-
The variables D, μ, DB, μB, DN, and μN are the quake catalogs. Geophys J Int 100:19–32
whole seismicity distribution, its mean inter-event Gardner JK, Knopoff L (1974) Is the sequence of earthquakes in
Southern California, with aftershocks removed, Poissonian?
time, the conditional distribution, and the mean inter-
Bull Seismol Soc Am 64(5):1363–1367
event time of the background and cluster series of Hainzl S, Scherbaum F, Beauval C (2006) Estimating back-
events, respectively. Finally, the following analytical ground activity based on inter-event–time distribution.
expression of the distribution D is obtained: Bull Seismol Soc Am 96(1):313–320
 Hutton K, Woessner J, Hauksson E (2010) Earthquake
μ μ monitoring in southern California for seventy-seven
D ð t Þ ¼ 1 ð1  n Þð1  D N ðt ÞÞ þ n ð1  DB ðt ÞÞ years (1932–2008). Bull Seismol Soc Am 100(2):423–
μN μB
446
ð47Þ Kagan Y, Knopoff L (1976) Statistical search for non-random
features of the seismicity of strong earthquakes. Phys Earth
Equation 20 then follows by derivation. Planet Inter 12:291–318
Knopoff L (2000) The magnitude distribution of declustered
earthquakes in southern California. Proc Natl Acad Sci U
References S A 97(22):11880–11884
Lawson AB (1993) A deviance residual for heterogeneous spa-
tial point processes. Biometrics 49:889–897
Baddeley A, Turner R, Moller J, Hazelton M (2005) Residual Luen B, Stark PB (2012) Poisson tests of declustered catalogs.
analysis for spatial point processes. J R Stat Soc Ser B Stat Geophys J Int 189(1):691–700
Methodol 67:617–666 Marsan D, Lengline O (2008) Extending earthquakes’ reach
Baddeley A, Moller J, Pakes AG (2008) Properties of residuals through cascading. Science 319:1076. doi:10.1126/
for spatial point processes. Ann Inst Stat Math 60:627–649 science.1148783
Baddeley A, Rubak E, Moller J (2011) Score, pseudo-score and Matthews MV, Ellsworth WL, Reasenberg PA (2002) A
residual diagnostics for spatial point process models. Stat Brownian model for recurrent earthquakes. Bull Seismol
Sci 26(4):613–646 Soc Am 92(6):2233–2250
Bak P, Christensen K, Danon L, Scanlon T (2002) Unified Molchan G (2005) Interevent time distribution in seismicity: a
scaling law for earthquakes. Phys Rev Lett 88:178501. theoretical approach. Pure Appl Geophys 162:1135–1150.
doi:10.1103/PhysRevLett.88.178501 doi:10.1007/s00024-004-2664-5
Beauval C, Hainzl S, Scherbaum F (2006) Probabilistic seismic Nanjo KZ, Ishibe T, Tsuruoka H, Schorlemmer D, Ishigaki Y,
hazard estimation in low seismicity regions considering Hirata N (2010) Analysis of completeness magnitude and
non-Poissonian seismic occurrence. Geophys J Int seismic network coverage of Japan. Bull Seismol Soc Am
164:543–550. doi:10.1111/j.1365-246X.2006.02863.x 100(6):3261–3268
Bottiglieri M, Lippiello E, Godano C, Arcangelis LDE (2009) Naylor M, Main IG, Touati S (2009) Quantifying uncertainty in
Identification and spatiotemporal organization of aftershocks. mean earthquake inter-event times for a finite sample. J
J Geophys Res 114(B03303):1978–2012 Geophys Res 114, B01316. doi:10.1029/2008JB005870
Christensen K, Danon L, Scanlon T, Bak P (2002) Unified scaling Newman W, Turcotte DL, Shcherbakov R, Rundle JB (2005)
law for earthquakes. Proc Natl Acad Sci U S A 99:2509–2513 Why Weibull? In: Abstracts of the American Geophysical
Corral A (2003) Local distributions and rate fluctuations in a Union Fall Meeting, San Francisco, California, 5−9
unified scaling law for earthquakes. Phys Rev E 68:035102. December 2005
doi:10.1103/PhysRevE.68.035102 Ogata Y (1988) Statistical models for earthquakes occurrences
Corral A (2004) Long-term clustering, scaling, and universality and residual analysis for point processes. J Am Stat Assoc
in the temporal occurrence of earthquakes. Phys Rev Lett 83(401):9–27
92:108501. doi:10.1103/PhysRevLett.92.108501 Omori F (1894) On the after-shocks of earthquakes. J Coll Sci
Corral A (2005) Mixing of rescaled data and Bayesian Tokyo Imp Univ 7:111–200
inference for earthquake recurrence times. Nonlinear Reasenberg P (1985) Second-order moment of Central
Process Geophys 12:89–100 California seismicity, 1969−1982. J Geophys Res 90:5479–
Corral A (2007) Statistical features of earthquake temporal occur- 5495
rence. In Bhattacharyya P, Chakrabarti BK (eds) Lect Notes Ripley BD (1976) The second-order analysis of stationary point
Phys, vol 705. Springer, Berlin, pp 191–221 processes. J Appl Probab 13:255–266
J Seismol (2013) 17:1041–1061 1061

Saichev A, Sornette D (2007) Theory of earthquake recurrence their interrelations. J Fac Sci Hokkaido Univ Ser VII
times. J Geophys Res 112(B04313):1–26. doi:10.1029/ 3:129–195
2006JB004536 Utsu T (2002) Statistical features of seismicity. Int Handb
Shcherbakov R, Yakovlev G, Turcotte DL, Rundle JB (2005) A Earthq Eng Seismol 81A:719–732
model for the distribution of aftershock waiting times. Phys Utsu T, Ogata Y, Matsu’ura RS (1995) The centenary of the
Rev Lett 95:1–4. doi:10.1103/PhysRevLett.95.218501 Omori formula for a decay law of aftershocks activity. J
Shcherbakov R, Turcotte DL, Rundle JB (2006) Scaling prop- Phys Earth 43:1–33
erties of the Parkfield aftershock sequence. Bull Seismol van Stiphout T, Zhuang J, Marsan D (2012) Seismicity
Soc Am 96(4B):S376–S384. doi:10.1785/0120050815 declustering. Community Online Resource for Statistical
Stoyan D, Grabarnik P (1991) Second-order characteristics for Seismicity Analysis. doi:10.5078/corssa-52382934. http://
stochastic structures connected with Gibbs point processes. www.corssa.org
Mathematische Nachrichten 151:95–100 Veen A, Schoenberg FP (2006) Assessing spatial point process
Talbi A (2009) Fluctuation of power law parameters in earth- models using weighted K-functions: analysis of California
quake inter-event time distribution and development of earthquakes. In: Baddeley A et al (ed) Lecture Notes in
fitting model. PhD thesis, Chiba University Statistics, vol 185. Springer, Berlin, pp 293–306
Talbi A, Yamazaki F (2009) Sensitivity analysis of the param- Weibull W (1951) A statistical distribution of wide applicability.
eters of earthquake recurrence time power law scaling. J J Appl Mech 18(3):293–297
Seismol 13:53–72. doi:10.1007/s10950-008-9115-1 Yakovlev G, Rundle JB, Shcherbakov R, Turcotte DL (2005)
Talbi A, Yamazaki F (2010) A mixed model for earthquake Inter-arrival time distribution for the non-homogeneous
inter-event times. J Seismol 14:289–307. doi:10.1007/ Poisson process. arXiv:cond-mat/0507657v1
s10950-009-9166-y Yakovlev G, Turcotte DL, Rundle JB, Rundle PB (2006)
Tibi R, Blanco J, Fatehi A (2011) An alternative and efficient Simulation-based distributions of earthquake recurrence
cluster-link approach for declustering of earthquake cata- times on the San Andreas fault system. Bull Seismol Soc
logs. Seismol Res Lett 82(4):509–518. doi:10.1785/ Am 96(6):1995–2007
gssrl.82.4.509 Yamanaka Y, Shimazaki K (1990) Scaling relationship between
Touati S, Naylor M, Main IG (2009) Origin and nonuniversality the number of aftershocks and the size of the main shock. J
of the earthquake inter-event time distribution. Phys Rev Phys Earth 38:305–324
Lett 102:168501. doi:10.1103/PhysRevLett.102.168501 Zhuang J (2006) Second order residual analysis of spatio-
Touati S, Naylor M, Main IG, Christie M (2011) Masking of temporal point processes and applications in model evalu-
earthquake triggering behavior by a high background rate ation. J R Statist Soc Series B 68(4):635–653
and implications for epidemic-type aftershock sequence in- Zhuang J, Ogata Y, Vere-Jones D (2002) Stochastic declustering
versions. J Geophys Res 116(B03304):1–26. doi:10.1029/ of space–time earthquake occurrences. J Am Stat Assoc
2010JB007544 97:369–380
Turcotte DL, Abaimov SG, Shcherbakov R, Rundle JB (2007) Zhuang J, Ogata Y, Vere-Jones D (2004) Analyzing earthquake
Nonlinear dynamics of natural hazards. In: Tsonis, features by using stochastic reconstruction. J Geophys Res
Anastasios A, Elsner, James B (eds) Nonlinear dynamics in 109, B05301. doi:10.1029/2003JB002879
geosciences. Springer, New York, pp 557–580. doi:10.1007/ Zoller G, Hainzl S (2007) Recurrence time distributions of large
978-0-387-34918-3_30 earthquakes in a stochastic model for coupled fault sys-
Utsu T (1969) Aftershock and earthquake statistics (I): Some tems: the role of fault interaction. Bull Seismol Soc Am
parameters which characterize an aftershock sequence and 97(5):1679–1687

View publication stats

You might also like