Zaliapin 2016

Geophysical Journal International
Geophys. J. Int. (2016) 207, 608–634 doi: 10.1093/gji/ggw300

Advance Access publication 2016 August 8
GJI Seismology
A global classification and characterization of earthquake clusters
Ilya Zaliapin1 and Yehuda Ben-Zion2

1 Department of Mathematics and Statistics, University of Nevada Reno, 1664 N Virginia St, Reno, NV 89557, USA. E-mail: zal@unr.edu
2 Department of Earth Sciences, University of Southern California, Los Angeles, CA, USA
Accepted 2016 August 4. Received 2016 August 3; in original form 2016 March 27
SUMMARY
Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

We document space-dependent clustering properties of earthquakes with m ≥ 4 in the 1975–
2015 worldwide seismic catalogue of the Northern California Earthquake Data Center. Earth-
quake clusters are identified using a nearest-neighbour distance in time–space–magnitude
domain. Multiple cluster characteristics are compared with the heat flow level and type of
deformation defined by parameters of the strain rate tensor. The analysis suggests that the
dominant type of seismicity clusters in a region depends strongly on the heat flow, while the
deformation style and intensity play a secondary role. The results show that there are two
dominant types of global clustering: burst-like clusters that represent brittle fracture in rela-
tively cold lithosphere (e.g. shallow events in subduction zones) and swarm-like clusters that
represent brittle–ductile deformation in relatively hot lithosphere (e.g. mid-oceanic ridges).
The global results are consistent with theoretical expectations and previous analyses of earth-
quake clustering in southern California based on higher quality catalogues. The observed
region-specific deviations from average universal description of seismicity provide important
constraints on the physics governing earthquakes and can be used to improve local seismic
hazard assessments.
Key words: Earthquake dynamics; Earthquake interaction, forecasting, and prediction;
Statistical seismology.
alternative approach discussed by Ben-Zion (2008, Sections 2 and

1 I N T RO D U C T I O N
3) considers the universal statistical descriptions to result in part
Seismicity is often discussed as a prime natural example of univer- from averaging data of large spatial domains having different event
sal self-similar behaviour (Bak & Tang 1989; Sornette & Sornette populations. If correct, clarifying the existence of different event
1989; Keilis-Borok 1990; Rundle et al. 2003; Corral 2004; Turcotte populations in relation to physical properties of fault zones and the
& Malamud 2004). The term ‘universality’ implies validity of the crust can increase the ability to extract detailed (region-specific)
same statistical laws in diverse geographic, geological, tectonic and information from observed data and improve the understanding of
physical settings; while ‘self-similarity’ refers to the abundance of earthquake physics.
earthquake characteristics described by power laws. A related term With these goals in mind, Bailey et al. (2009, 2010) analysed pat-
to ‘self-similarity’ is ‘scale-invariance’. We recall that the only func- terns of earthquake focal mechanisms in southern California and
tion that is invariant with respect to changes of measurement units found persisting differences in relation to geometrical properties of
and/or scale of analysis is a power law. Table 2 in Ben-Zion (2008) the major fault zones. Zaliapin & Ben-Zion (2011) analysed along-
lists various examples of power-law distributions of earthquake and strike symmetry properties of aftershocks in catalogues of 25 fault
fault quantities. The most established of those are the power-law zones in California, and established relations between deviations
distribution of seismic moments (Kagan 1999), which is an alterna- from generic symmetric distribution and contrasts of seismic veloc-
tive form of the exponential distribution of earthquake magnitudes ities across the faults. See also Rubin & Gillard (2000) and Rubin
in the Gutenberg–Richter law (Gutenberg & Richter 1944), and the (2002). The results are consistent with theoretical expectations on
power decay rate of events following a large earthquake referred to differences between ruptures on faults that do or do not separate
as the Omori–Utsu law (Omori 1894; Utsu et al. 1995). different elastic solids (e.g. Weertman 1980; Ben-Zion 2001; Am-
These laws were claimed to be universal on a global scale, at least puero & Ben-Zion 2008). Yang & Ben-Zion (2009) and Enescu et
at geologically long time intervals (e.g. Kagan 1999). In this view al. (2009) showed that parameters of the Omori–Utsu aftershock
the documented discrepancies in observed forms and parameters decay law in southern California are correlated with the heat flow.
of earthquake statistics are attributed to statistical fluctuations and Zaliapin & Ben-Zion (2013a,b) took this further by showing that
artefacts of catalogue uncertainties (e.g. Kagan 1999, p. 569). An there are distinctly different types of seismicity clusters in southern
608
C The Authors 2016. Published by Oxford University Press on behalf of The Royal Astronomical Society.
Global analysis of earthquake clusters 609
Table 1. Statistics of singles, main shocks, aftershocks and foreshocks in the cluster analysis of events with m ≥ 4 in the NCEDC catalogue during
1975–2015.
Magnitude range Singles Families
Main shocks
(= no. of families) Aftershocks Foreshocks
No. Per cent No. Per cent No. Per cent No. Per cent
All events: m ≥ 4 116 228 45.2 19 612 7.6 105 790 41.2 15 363 6.0
4≤m<5 98 385 47.9 8098 3.9 86 749 42.2 12 197 5.9
5≤m<6 17 503 36.8 9141 19.2 17 999 37.8 2968 6.2
6≤m<7 340 9.6 2049 57.7 973 27.4 187 5.3
7≤m<8 0 0 298 79.5 66 17.6 11 2.9
m≥8 0 0 26 89.7 3 10.3 0 0
California with preferred locations correlated with the heat flow. making them more swarm-like, and moves some cluster events to

These results are consistent with theoretical expectations on prop- the background mode. Accordingly, working with low-quality cat-
erties of earthquake sequences in regions with different effective alogues requires developing statistics tools that are robust to the
viscosity (Ben-Zion & Lyakhovsky 2006). In the present paper we catalogue uncertainties.
generalize the results of Zaliapin & Ben-Zion (2013a,b) to the global In the following sections we develop such a toolbox, and use
scale. it to reveal strong spatial dependence of global earthquake clus-
Our analysis focuses on earthquake clustering—partitioning of tering that is mainly controlled by the local heat flow production.
seismicity into groups closer in space and time than expected in We confirm the existence of the two primary types of earthquake
a purely random distribution. Such groups reflect diverse trigger- clusters—burst-like and swarm-like—and show that burst-like clus-
ing processes and prominently include traditional aftershock series, ters are associated with cold regions (mainly shallow seismicity of
but also swarms and other types of clustering (Zaliapin & Ben- subduction zones), while swarm-like clustering is typical for hot
Zion 2013a; Vidale & Shearer 2006; Vidale et al. 2006; Zhang & regions (mainly mid-oceanic ridges). The type of plate-boundary
Shearer 2016). Facilitated by high-quality catalogue and problem- deformation is also examined and shown to play a secondary role
specific statistical techniques, we demonstrated in an earlier study in determining the cluster style of seismicity. The global results
that the cluster style of seismicity in southern California is closely presented in this study are consistent with our previous regional
related to physical properties of the crust and is changing at the findings in southern California based on higher-quality data. The
scale of tens of kilometres (Zaliapin & Ben-Zion 2013b). In par- analysis of possible sources of artefacts for each examined statis-
ticular, it was shown that there are two dominant types of clusters: tics provide results designed to be robust to the known catalogue
(i) ‘Burst-like clusters’ with a prominently large main shock, small uncertainties and deficiencies.
number of foreshocks and dominance of first-generation offspring.
Such clusters reflect highly brittle rapid failure process in areas with
cold crystalline rocks, decreased fluid content, and low heat flow 2 D ATA A N D M E T H O D S
production (increased effective viscosity). Burst-like cluster areas
in southern California include the San Jacinto fault zone, Mojave, 2.1 Earthquakes
Ventura and San Gabriel regions. (ii) ‘Swarm-like clusters’ that We work with the global catalogue produced by the Northern Cali-
lack a prominent main shock, have increased foreshock activity, fornia Earthquake Data Center (NCEDC 2015). The examined cat-
and abundance of secondary, tertiary, etc. offspring. Such clusters alogue covers the period 1/1/1975 to 6/9/2015 and contains 256 993
reflect mixed brittle–ductile failure in areas with increased fluids events. The minimal magnitude used in the analysis is mmin = 4.
and heat flow and/or soft sediments (decreased effective viscosity). This magnitude is higher than the completeness magnitude in many
Swarm-like cluster areas in southern California include the Salton examined regions, in particular during the earlier times. We demon-
Sea and Coso geothermal regions. The quality of data in south- strated (Zaliapin & Ben-Zion 2013a, Appendices D and E) that the
ern California allowed us to validate the region-specific character cluster structure estimated by our technique (Section 3) is insen-
of earthquake clustering by statistical differences in thirteen com- sitive to the catalogue incompleteness as well as to the minimal
plementary cluster characteristics, including aftershock/foreshock reported magnitude. Accordingly, some cluster statistics, like the
intensity, magnitude difference between main shock and the largest total number of clusters and partition of events into main shocks,
aftershock/foreshocks, b-value, cluster area, duration, etc., all of foreshocks, and aftershocks (see Section 3.6 for definitions) are
which related to the effective viscosity of a region and hence to the fairly robust with respect to the magnitude incompleteness. The
cluster type (Zaliapin & Ben-Zion 2013b, Table 1 and Appendix C). incompleteness however does affect the cluster size distribution, as
The above results from southern California demonstrate the exis- discussed in Section 4.2.
tence of region-specific features that provide important information We only consider events with depth less than zc = 70 km. The
on earthquake dynamics and can contribute to improving seismic depth reporting in NCEDC catalogue is highly irresolute: 69 932
hazard assessments. However, extending the results to the global events (27.2 per cent) are assigned a depth of 33 km and 61 939
scale faces the problem of data quality. This is because higher mag- events (24.1 per cent) are assigned a depth of 10 km. Other popular
nitudes of completeness/reporting and earthquake location uncer- (default) depth values are 35 km (14 421 events, 5.6 per cent), 30 km
tainties impact cluster identification and lead to multiple artefacts (4899 events, 1.9 per cent), and 5 km (2651 events, 1.0 per cent).
(Zaliapin & Ben-Zion 2015). In particular, low catalogue quality In addition, there is a tendency, especially during earlier times, to
blurs the underlying fine structure of earthquake clusters, artificially assign depths divisible by 5 km (5, 10, 15, etc.). Our analysis is
610 I. Zaliapin and Y. Ben-Zion
based on earthquake epicentres and is not affected by the depth Here ei are the eigenvalues of the strain rate tensor. The strain
uncertainties. rate tensor style S can be used to roughly quantify the type of
Fig. 1(a) shows the spatial intensity (x) of events in the NCEDC displacement into contraction (S < −0.5), strike-slip (0.5 < S <
catalogue, in events with magnitude m ≥ 4 per year per 10 000 −0.5), and extension (S > 0.5). The maps of strain rate tensor
km2 . Appendix B describes the process of producing smooth spatial second invariant and style are shown in Fig. A2.
maps of different seismic and physical characteristics used in this
study. The intensity varies over several orders of magnitude, from
0.02 to 12.5, with the highest values associated with contracting 2.4 -Analysis
subduction zones and lowest values associated with mid-oceanic
spreading ridges. The global spatial distribution of the maximal Any cluster analysis of earthquakes is affected by the existence of
observed earthquake magnitude mmax of the examined seismicity the catalogue lower cut-off magnitude mmin (which may be smaller
is illustrated in Fig. 1(b). Naturally, the fluctuations of the maxi- than the completeness magnitude mc ). For instance, if we analyse
mal observed magnitude are closely related to the seismic intensity earthquakes with m ≥ mmin = 4, then an earthquake of m = 4 cannot
fluctuations. The distribution of hypocentral depth of the examined have recorded aftershocks of a smaller magnitude, while an m = 6
earthquakes is shown in Fig. A3(a). Its spatial variations resemble event may have aftershocks with magnitudes 4 ≤ m ≤ 6. To equalize
those for the earthquake intensity and maximal magnitude. the magnitude ranges for potential fore/aftershocks of main shocks
with different magnitudes, we sometimes perform -analysis that

Fig. 1(c) presents the spatial variability of the magnitude com-
pleteness of the examined catalogue. It shows the proportion p5 of (i) only considers main shocks with magnitude m ≥ mmin + and
earthquakes with magnitude m ≤ 5, which serves as a proxy to com- (ii) only considers fore/aftershocks with magnitude within units
pleteness quality. Specifically, if the number N(m) of events with below that of a main shock. The fore/aftershocks detected by this
magnitude above m is given by the Gutenberg–Richter law analysis are called -fore/aftershocks. The conventional analysis
that considers all events is referred to as regular analysis.
log10 N (m) = a − bm, m ≥ m c (1)
with b = 1 and mc = 4, then p5 = 0.9. Deficiency of events in 3 E A RT H Q UA K E C L U S T E R S

small magnitude range results in lower values of p5 . Mild fluctu-
ations of the b-value also affect this completeness proxy albeit to 3.1 Generalized earthquake distance
a lesser degree. For instance, if we assume validity of eq. (1) with
mc = 4, then variation of b-values in the range [0.8, 1.2] results in Consider a catalogue where each event i is characterized by its
p5 in the range [0.84, 0.94], with lower b-values corresponding to occurrence time ti , hypocentre (φ ι , λi, di ), and magnitude mi . We
lower p5 values. Hence, the main fluctuations of p5 (that goes as define the proximity ηij of earthquake j to earthquake i following
low as 0.5) observed in Fig. 1(c) are primarily due to incomplete- Baiesi & Paczuski (2004) as:
ness. The analysis suggests that catalogue quality is deteriorated
ti j (ri j )d 10−bm i , ti j > 0;
for oceanic seismicity, relatively far from seismic networks, which ηi j = (4)
mainly occur in the southern hemisphere. ∞, ti j ≤ 0.
Fig. 1 illustrates and summarizes the diversity of seismic regimes
Here, tij = tj − ti is the event intercurrence time, which is positive
and parameters as well as variations in catalogue quality involved
if event j occurred after event i; rij ≥ 0 is the spatial distance between
in a global study. The statistics used in our analysis are designed
the earthquake hypocentres; d is the (possibly fractal) dimension
to be robust with respect to these obstacles and yet still reflect the
of the hypocentres or epicentres, and b is the parameter of the
essential characteristics of the regional cluster style.
Gutenberg–Richter law (1). The motivation for and properties of this
proximity measure are discussed in Zaliapin & Ben-Zion (2013a,
2015, 2016). In particular, the proximity to the nearest neighbour
2.2 Heat flow is inversely related to the seismic intensity. It is intuitive that the
The employed surface heat flow data is taken from Bird et distance between events is smaller in a high-intensity process where
al. (2008). The heat flow within the seismically active areas is a larger number of events occupy the same space–time volume; see
mapped in Fig. 2. The distribution of the heat flow over the en- Zaliapin et al. (2008) for formal derivations and Zaliapin & Ben-
tire Earth surface is shown in Fig. A1. The heat flow production Zion (2013a) for simulation results.
is prominently high along the oceanic spreading ridges, reaching
0.3 W m−2 .
3.2 Parent-offspring identification
For each event i we identify its unique nearest neighbour (parent)
2.3 Strain rate tensor j with respect to the distance given by eq. (4), and denote for sim-
plicity the nearest-neighbour distance by the same symbol ηij . The
We use the global strain rate field modelled by Kreemer et al. (2014). event i is called an offspring of j. According to this definition, each
Specifically, we consider the second invariant I2 of the strain rate event (except the first one in the catalogue) has a unique parent, and
tensor ε̇: also might have multiple offspring.

2 2
I2 = ε̇ϕϕ + (ε̇θθ )2 + 2 ε̇ϕθ (2)
and the tensor style S defined by Kreemer et al. (2014) as: 3.3 Bimodal distribution of the nearest-neighbour distance
e1 + e2 Consider the space and time distances between event i and its par-
S= . (3)
max (|e1 | , |e2 |) ent j normalized by the magnitude of the parent event (Zaliapin
Figure 1. Spatial distribution of selected statistical characteristics of earthquakes with magnitude m ≥ 4 according to the global NCEDC catalogue during
1975–2015. A point is included in this graph if the circle of radius 100 km centred at the point contains 5 or more earthquakes of magnitude m ≥ 5. Red lines
depict major tectonic faults. Shades correspond to bathymetry and topography. Continents are depicted by grey colour. (a) Earthquake intensity in events ×
yr−1 × 10 000 km−2 (b) Maximal observed magnitude mmax . (c) Proportion of events with m ≤ 5.

Figure 2. Heat flow in the seismically active regions. A point is included in this graph if the circle of radius 100 km centred at the point contains 5 or more
earthquakes of magnitude m ≥ 5. The heat flow values are clipped at H = 0.2 (the maximal reported value is 0.3).
et al. 2008): give very close results. The details of numerical implementation
are discussed in Hicks (2011). A model assigns to each event the
Ti j = ti j 10−qbm i ; Ri j = (ri j )d 10− pbm i ; q + p = 1. (5) probabilities w and (1−w) of being attributed to one or the other
This is convenient because now log ηi j = log Ti j + log Ri j . mode. We make the final mode assignment according to the maximal
Zaliapin et al. (2008) and Zaliapin & Ben-Zion (2013a) demon- probability v = max (w, 1−w). This corresponds to choosing the
strated that a time-stationary space-inhomogeneous Poisson flow of mode separation threshold η0 that equalizes the densities of the two
events with Gutenberg–Richter magnitudes corresponds to a uni- estimated Gaussian modes:
modal distribution of (log T, log R) that is concentrated along a line
N (η0 ; μ1 , 1) = N (η0 ; μ2 , 2 ). (7)
log10 T + log10 R = const. Observed seismicity, however, shows a
bimodal joint distribution of (log10 T, log10 R), as has been docu- The background (cluster) events now can be equivalently defined
mented in multiple studies of various regions (e.g. Zaliapin et al. by the condition ηij > η0 (ηij ≤ η0 ). In 1-D case, the position ηbg
2008; Zaliapin & Ben-Zion 2011, 2013a,b, 2015, 2016; Gu et al. of the background is defined as the mean value of the estimated
2013; Davidsen et al. 2015; Reverso et al. 2015; Schoenball et al. rightmost Gaussian mode: ηbg = max (μ1 , μ2 ). In 2-D case we
2015). One of the modes is similar to that observed in a Poisson pro- define ηbg = max (μ1 [1]+μ1 [2], μ2 [1]+μ2 [2]), where the notation
cess and corresponds to background events, while the other consists [i] refers to the ith component of a vector. Alternatively, one can
of clustered events located considerably closer in time and space to define ηbg as (a) the mean generalized earthquake distance ηij of
their parents than expected in a Poisson process (see Fig. 4). the background events, or (b) the mean generalized distance of
events that happened at large spatial distance from their parent (say,
R > 5, 50). The last approach is motivated by the observation that the
3.4 Separating the background and cluster modes: a majority of events at large spatial distances from their parent belong
Gaussian mixture model approach to the background mode (see Fig. 4). These alternate approaches
give results (not shown) that are very close to those obtained with
Analysis of statistical properties of the background and cluster
our main method.
modes is one of the main tools of this study (see Section 4.2).
The regional mode separation quality Q is defined as the average
The bimodality of the earthquake distance distribution (Figs 4b and
value of the mode assignment probability v = max (w, 1 − w)
d) allows one to use a suitably chosen nearest-neighbour threshold
over all events in a region. According to this definition, the quality
η0 to formally attribute each event to either the background (if ηij
is constrained by 0.5 ≤ Q ≤ 1, where Q = 1 corresponds to a
> η0 ) or cluster (if ηij < η0 ) mode. The threshold selection is done
perfect separation (each event is attributed to one of the modes
here according to a Gaussian mixture model with two modes.
with probability 1) and Q = 0.5 corresponds to an indeterminate
A two-mode Gaussian mixture model assumes that sample xi
separation (each event is attributed to either mode with the same
Rm , i = 1, . . . , n comes from the distribution
probability of 0.5).
F(x) = wN (x; μ1 , 1) + (1 − w)N (x; μ2 , 2) , (6)
where w is the mixture weight of the first mode and N(x;μ, )
3.5 Cluster identification
denotes the Gaussian (Normal) distribution, with mean μ that is a
vector with m components and variance that is a positive-definite Connecting each earthquake in the catalogue to its nearest neigh-
m × m matrix. The estimation of such model can be done using the bour (parent) according to the nearest-neighbour distance η of eq.
Expectation-Maximization algorithm (Dempster et al. 1977). (4) produces a single cluster (spanning network) that contains all
In our setting, we can either apply a 1-D Gaussian mixture model examined events. From a graph-theoretical perspective, this cluster
to the log-distance log10 η or a 2-D Gaussian mixture model to is a tree graph, which means that it does not have loops (Zaliapin
the logarithmic components (log10 T, log10 R). Both approaches & Ben-Zion 2013a; Baiesi & Paczuski 2004). Removing all links
that correspond to large parent-offspring distances, defined by the Fig. 3(b) shows the average value of seismic intensity for events
condition η ≥ η0 , creates a spanning forest—a collection of trees with m ≥ 4 in the same coordinates (S, I2 ). The highest average
each representing a separate earthquake cluster. The forest contains intensity of ≈ 3 events per year per 10 000 km2 is exclusively
many single-event trees, which we call singles. The other clusters observed within contracting environments: S < −0.75, I2 > 100.
contain multiple events and are called families. The lowest average intensity, 0.2 < < 1, is observed within
extending environments. The intensity in strike-slip zones is in-
termediate around and slightly below = 1. The highest seismic
3.6 Event classification activity typically occurred in subduction zones, which explain a
In each family, the earthquake with the largest magnitude is called close resemblance in the patterns of seismic intensity (Fig. 3b) and
main shock. If there are several earthquakes with same largest mag- that of the average hypocentral depth (Fig. A4a). The average heat
nitude within a family, the first one is considered to be the main flow has the highest values (H > 0.25) exclusively within extension
shock, so each family has a single main shock. All events in a fam- environments—along the mid-oceanic spreading ridges (Fig. 3c).
ily that occurred after the main shock are called aftershocks. All The results in Figs 3(b) and (c) emphasize that the spatial distri-
events that occurred prior to the main shock are called foreshocks. bution of seismic intensity is inversely related to that of the heat
Each single is also considered to be a main shock (that has no flow. We also perform Spearman’s rank correlation analysis (see
foreshocks and aftershocks). Appendix C for definition and discussion) and generalized linear

model (GLM) analysis (see Appendix D) for earthquake intensity
versus heat flow H and strain rate tensor parameters S and I2 . The
3.7 Parameters results, summarized in Tables 5, 6 and Figs D1(a), (b) and (c), cor-
roborate the observations of Fig. 3. We show below that the heat flow
In this study we use event epicentres rather than hypocentres be- production also governs the space-dependent style of earthquake
cause the depth coordinates are often less accurate than those of the clustering.
epicentres, and location errors lead to various analysis artefacts as Fig. 4 presents the distribution of the generalized earthquake
discussed in detail by Zaliapin & Ben-Zion (2015). In addition, we distance η of eq. (4) and the joint distribution of its normalized space
fix b = 1, d = 1.3 and p = 0.5 in eqs (4) and (5). Zaliapin & Ben- and time components (R, T) of eq. (5) for the earthquakes from areas
Zion (2013a) showed that the estimated cluster structure is fairly with low (H < 0.1) and high (H > 0.25) heat flow value. Panels
robust with respect to the values of these parameters. Accordingly, (a, c) show the joint two-dimensional density of (R, T) estimated for
the main conclusions of this study are not sensitive to the precise all earthquakes in areas with the indicated heat flow level. Colours
parameter values. represent the relative number of points with given values of (R, T),
We refer to Zaliapin & Ben-Zion (2013a) for further details on as indicated in the colourbar on the right. The integral over the entire
and examples of performance of our cluster technique, as well as distribution is 1. Panels (b, d) show the estimated density for the
detailed analysis on its stability. The statistical artefacts of catalogue nearest-neighbour distance η on a logarithmic scale—the histogram
uncertainties that affect cluster analysis based on parent-offspring values for log10 η divided by the total number of examined events—
identification are examined in Zaliapin & Ben-Zion (2015). so that the integral over the distribution is 1. The background and
cluster modes are seen clearly in both cases. However, there exist
several notable differences in the cluster style of earthquakes in
4 R E S U LT S
high versus low heat flow regions: (i) The typical position ηbg of
the background mode in low heat flow regions (log10 ηbg ≈ −4.7)
4.1 Basic characteristics of seismicity
is lower than that in high heat flow regions (log10 ηbg ≈ −4.0). This
For the purpose of this study, a point on the Earth surface is denoted reflects higher earthquake intensity in low heat flow regions that
seismically active if there are 5 or more events with magnitude m ≥ have predominantly contracting and transform deformation style,
5 within 100 km of this point according to the NCEDC catalogue and corroborates our earlier observations in Figs 3(b) and (c). (ii)
during 1975–2015. The total Earth surface area is 510.1 × 106 km2 ; The proportion of background events in low heat flow regions is
about 9 per cent of this area, or 44.9 × 106 km2 , is seismically active. lower than that in high heat flow regions. Accordingly, the proportion
Fig. 1 shows basic characteristics of seismicity in the active areas: of clustered events is higher in low heat flow regions. (iii) The
(panel a) Intensity, in event/(year × 10 000 km2 ), of earthquakes time decay of cluster events is faster in high heat flow regions,
with magnitude m ≥ 4, (panel b) maximal observed magnitude and leading to stronger time separation between the background and
(panel c) proportion of events with magnitude m ≤ 5 which serves as cluster modes. This can be seen by comparing how the cluster
a proxy for catalogue completeness (see discussion in Section 2.1). mode is blending with the background mode in panels (a) and
Next we relate the three characteristics illustrated in Fig. 1 to the (c) of Fig. 4. The offspring duration is longer in low heat flow
type of lithospheric deformation. Fig. 3(a) displays the partitioning regions, which is reflected by a horizontally elongated shape of the
of seismically active area with respect to values of the strain rate cluster mode in Fig. 4(a), as opposed to a more confined location
tensor style S and second invariant I2 . The analysis is done within of the cluster mode in high heat flow regions in Fig. 4(c). (iv)
0.5◦ × 0.5◦ cells that tile the Earth surface. About half of the Proportion of repeaters—events that happen at short spatial and
active area corresponds to lowest values of the second invariant large temporal distances from the parent and hence occupy the
(I2 < 100) attributed to subduction zones and other contraction lower right corner of the (T, R) plots in Figs 4(a) and (c)—is larger
environments. The largest 10 per cent of the values of the second in high heat flow regions. This observation is further illustrated
invariant (I2 > 1000) are associated with mid-ocean ridges and in Fig. 5 that shows the distribution of rescaled time to parent T
other extension environments. The partition of seismic area into for offspring within two parent rupture lengths to the parent. In
different tensor styles is fairly independent of the values of I2 — cold regions (panel a) the cluster and background modes are largely
approximately 60 per cent in contraction, 20 per cent in strike-slip overlapping at these short distances to parent. The background mode
and 20 per cent in extension. is centred at about log10 T = −3 and has lower intensity and spread
Figure 3. Average values of selected statistics as a function of strain rate tensor style S and second invariant I2. (a) Seismogenic area A [km2 ] (the values are
reported on a logarithmic scale). (b) Seismic intensity [event yr–1 10 000 km–2 ] (the values are reported on a logarithmic scale). (c) Heat flow H [W m–2 ].
compared to that of the clustered mode. The latter is centred at is centred at about log10 T = −2 (lower intensity of background
log10 T = −6 and has much larger spread, interpreted as slow decay events compared to those in cold regions), and has notably higher
of intensity of offspring earthquakes. In hot regions (panel b), on the intensity than the cluster mode. The cluster mode is centred at about
contrary, the two modes are well separated. The background mode log10 T = −6.5 and has smaller spread than that of the cluster mode

Figure 4. Generalized earthquake distance η of eq. (4) and its normalized space and time components (T, R) of eq. (5) in regions with different level of heat
flow H. (a, b) Earthquakes in regions with low heat flow, H < 0.1. (c, d) Earthquake in regions with high heat flow, H > 0.25. (a, c) Join distribution of the
rescaled components (T, R) of the earthquake nearest-neighbour distance. (b, d) Distribution of the values of the nearest-neighbour distance η. Black diagonal
lines in panels (a, c) depict levels of constant distance η (from top to bottom): –log10 η = 4, 5, 6, 7, 8.
in cold regions, suggesting faster decay of intensity of offspring (panel b). The map of the quality Q of the mode separation is shown
earthquakes. in Fig. A3(c).
The location of the background mode is primarily controlled by
the absolute intensity of the background events (Zaliapin et al. 2008;
Zaliapin & Ben-Zion 2013a). This explains the inverse relation be-
tween the background location (Fig. 6a) and earthquake intensity
4.2 Cluster and background modes
(Fig. 1a). Furthermore, the values of ηbg follow a three-modal distri-
The results of Figs 3–5 demonstrate that earthquake clustering style bution, clearly outlining the major tectonic environments in agree-
is space-dependent and related to the heat flow production. We now ment with Fig. 3(b). The highest earthquake intensity and lowest
complement these analyses by additional statistics involving the values of ηbg < −4.5 are observed within convergent environments.
earthquake nearest-neighbour distances. Specifically, we apply a The lowest earthquake intensity and largest values of ηbg > −3.75
1-D Gaussian mixture model (Section 3.3) to the nearest-neighbour are observed along divergent boundaries. Intermediate values of
distances log10 η of events within circles of radius r = 200 km earthquake intensity and background position −4.5 < ηbg < −3.55
centred at the epicentres of all examined earthquakes. The model is are observed along transform boundaries. Fig. A3(b) shows the
used to estimate the space-dependent threshold η0 that separates the worldwide spatial distribution of a related feature—the threshold
cluster and background modes, partition the events into cluster and η0 that separates the background and cluster mode, according to a
background populations, estimate the characteristic position ηbg of Gaussian mixture model.
the background mode, and assess the quality Q of mode separation. The other examined cluster characteristics exhibit similar spatial
Fig. 6 shows the spatial maps of the position ηbg of the background variations. In particular, divergent environments have uniformly
mode (panel a) and the proportion of events in the background mode increased background proportions pbg > 0.7 (Fig. 6b) and high

Figure 5. Distribution of the rescaled time to parent T of eq. (5) for offspring that occurred within two parent rupture distances from the parent. (a) Cold
regions, H < 0.2. (b) Hot regions, H > 0.2.
mode separation quality Q ≈ 0.95 (Fig. A3c). Convergent envi- 4.3 Properties of earthquake clusters
ronments exhibit much larger spatial variability and intermittence
The 256 993 events of the examined catalogue have been partitioned
in the values of background proportions pbg and separation quality
into 135 840 clusters according to the procedure of Section 3. Of
Q. For instance, Fig. 6(b) shows that in the Northwestern part of
those clusters, 116 228 (85.6 per cent) are singles and 19 612 (14.4
the Pacific plate, along the Kuril-Kamchatka and Japan trenches,
per cent) are families with sizes ranging from 2 to 6584. Tables 1 and
the background proportion varies widely in the range 0.2–0.7 over
2 summarize the individual event classification (into singles, main
a scale of hundreds of kilometers that coincides with the spatial
shocks, foreshocks, and aftershocks) in the regular and -analysis,
resolution of our analysis. Another example of highly intermit-
respectively.
tent spatial behaviour is the Persia-Tibet-Burma orogeny in the
Fig. 8(a) shows the distribution of cluster size N for clusters in
Eurasian plate. Overall, however, the average proportion of the
areas with high (H > 0.2) and low (H < 0.2) heat flow levels. The
background events in transform and convergent environments is
distribution tail in both cases can be approximated by a power law
lower than in divergent environments, as illustrated in Fig. 7(b).
Similarly, the mode separation quality Q shows high intermittency S (N ) = Prob [cluster size > N ] ∝ N −α (8)
within transform and convergent boundaries. It changes in the range
between 0.9 and 0.95 with rather sharp spatial gradients (Fig. A3c), with power index α ≈ 2 in hot areas and α ≈ 1 in cold areas. The
and has overall lower values compared to divergent zones (Fig. value of α ≈1 was previously reported for the cluster size distri-
7c). The observed clear spatial variations in the cluster parame- bution in southern California (Zaliapin & Ben-Zion 2013a). The
ters are not spurious but governed by local tectonic and physical observed difference in the cluster size distributions implies that (i)
settings. This was shown in a local study for southern California cold areas have much larger clusters—indeed, the maximal cluster
(Zaliapin & Ben-Zion 2013b); a detailed demonstration of such size in cold areas is max (N | H < 0.2) = 6584 while the maximal
correlations in the global setting is outside the resolution of this cluster size in hot areas is 35 times smaller, max (N | H ≥ 0.2) =
study. Fig. 7 compares the three examined parameters of seismic 186; and (ii) the proportion of clusters with size N > 10 is larger
clustering as functions of strain rate tensor’s style S and second in cold areas. Recall that the cluster size statistically increases with
invariant I2 . The comparison of earthquake cluster statistics with the maximal observed magnitude, since larger events have more
heat flow and strain rate tensor parameters using Spearman’s cor- offspring (e.g. Utsu 1970); it also increases as the magnitude of
relation and GLM approach is illustrated in Tables 5, 6 and Figs completeness decreases. Accordingly, the dominance of large clus-
D1(d)–(o). This further documents the coupling between the exam- ters in cold regions observed in Fig. 8(a) is explained by statistically
ined cluster characteristics and their correlation with the heat flow higher maximal magnitude (Fig. 1b) and better quality of catalogues
(cf. Fig. 3c). (Fig. 1c) in cold regions compared to hot ones.

Figure 6. Global maps of selected parameters of seismic clustering. (a) Average nearest-neighbour distance log10 ηbg in the background mode. (b) Proportion
of events in the background mode.
To eliminate effects related to differences in the largest regional magnitude m > 6. At the same time, the size of intermediate-
magnitude, we compare the cluster size distributions in relatively magnitude clusters (with main shock magnitude m < 6) is stochas-
hot and cold areas using -analysis with = 2 (Fig. 8b). This tically larger in hot regions. These two observations are consistent
approach equalizes the cluster sizes with main shocks of different with the finding of Zaliapin & Ben-Zion (2013b) in southern Cal-
magnitudes (Zaliapin & Ben-Zion 2013a), and hence should elim- ifornia, who also pointed out the difference in clustering styles of
inate the discrepancy caused by different levels of seismic activity the largest regional events and the rest of earthquakes, and reported
in cold and hot regions. The results indicate that the cluster size is larger cluster size of intermediate-magnitude clusters in hot regions.
stochastically larger in cold area. Recall that a random variable X As a particular case of small-cluster size analysis, we notice that
is said to be stochastically larger than Y if the survival function of the proportion pS of smallest clusters—singles—among all detected
X is larger than that of Y for all arguments. Finally, we compare the clusters is higher in cold areas: pS (H < 0.2) = 0.86 vs pS (H ≥ 0.2)
cluster size distribution for clusters with intermediate-magnitude = 0.80. The observed difference in proportions is highly signif-
main shocks. As shown in Fig. 8(c), the size of clusters with main icant, with p-value being essentially zero (p < 10−16 ) according
shock magnitude m < 5 is stochastically larger in hot regions. to the Fisher test (Agresti 2007). This effect is noteworthy, since
Similarly, the cluster size is stochastically larger in hot regions for the higher maximal magnitude, better quality of catalogues, and
clusters in all magnitude ranges below m = 6 (not shown). In ad- lower completeness magnitude in cold regions should decrease the
dition, the number of foreshocks and aftershocks per cluster with number of singles (e.g. Zaliapin & Ben-Zion 2015). On the other
main shock magnitude m < 6 is significantly higher in hot regions hand, the probability of being a single is higher for small-magnitude
(not shown). events (e.g. it is more probable for m = 4 event to have no offspring
In summary, stochastically larger cluster size in cold regions is than for m = 7). Accordingly, an increased detected proportion
related to the presence of large-magnitude clusters with main shock of small-magnitude events in cold regions (Fig. 1c) might inflate
Figure 7. Statistics of background and cluster modes as functions of strain rate tensor’s style S and second invariant I2 . (a) Average nearest-neighbour distance
log10 ηbg in the background mode. (b) Proportion of background events pB . (c) Quality Q of background/cluster mode separation.
the proportion of singles. We demonstrate that this effect is not singles is higher in cold regions (not shown) for each magnitude
ultimately responsible for an increased proportion of singles in interval from [4.0, 4.1), to [5.5, 5.6). Table 3 summarizes the re-
cold regions by repeating the analysis within each magnitude in- sults of a formal statistical testing that uses magnitude intervals of
terval of length 0.1: [4.0, 4.1), [4.1, 4.2), etc. The proportion of length 0.5 and confirms the statistical significance of the observed
Table 2. Statistics of singles, main shocks, aftershocks and foreshocks in the cluster -analysis of events with m ≥ 2 in the NCEDC catalogue during
1975–2015.
Magnitude range Singles Families
Main shocks
( = no. of families) Aftershocks Foreshocks
No. Per cent No. Per cent No. Per cent No. Per cent
All events: m ≥ 4 404 1.9 2309 10.9 16 165 76.3 2314 10.9
4≤m<5 – – – – 9280 88.4 1,216 11.6
5≤m<6 – – – – 6067 86.9 913 13.1
6≤m<7 390 11.8 1999 60.3 750 22.6 174 5.2
7≤m<8 14 3.7 284 75.9 65 17.4 11 2.9
m≥8 0 0 26 89.7 3 10.3 0 0
(a) Regular analysis (b) Δ−analysis, Δ=2

0 0
10 10
High heat flow, H > 0.2 High heat flow, H > 0.2

−1 Low heat flow, H < 0.2 Low heat flow, H < 0.2
10
Survival function, S(N)

−1
−2 slo 10
10 pe
−1
−3
10
slo
sl
−2
pe
10
op
e
−2
−2
−4
10
−5
10 −3
10
−6
10 0 1 2 3 4 0 1 2
10 10 10 10 10 10 10 10
Cluster size, N Cluster size, N
0
(c) Regular analysis, mainshock m < 5
10
High heat flow (H>0.2)
Low heat flow (H<0.2)
−1
10
−2 slo
10 pe
−2
. 5
−3
10
−4
10
−5
10 0 1 2
10 10 10
Cluster size, N
Figure 8. Distribution of cluster size N in regions with high (H > 0.2, red solid line) and low (H < 0.2, blue dashed line) values of heat flow. For families, the
heat flow value is estimated at the main shock epicentre. The y-axis shows the survival function S(N) = Prob.[cluster size > N]. The lines that correspond to
power laws S(N) ∝ N−α with indices α = 1 and α = 2 are shown for visual convenience. (a) Regular analysis, all clusters. (b) Delta analysis with = 2. (c)
Regular analysis, clusters with main shock magnitude m < 5.
differences for events with magnitudes from 4.0 to 5.5. This analysis Fig. 9(a) shows the proportion pS of singles among families in
also demonstrates that events with m > 5.5 in cold regions become different regions. The proportion varies between 0.7 and 0.95. The
singles less often than those in hot regions (not shown). It is difficult highest values (pS > 0.9) are typically observed within cold regions,
to conclude with the existing data whether this effect is related to while lowest values (pS < 0.75) almost exclusively belong to hot
the inferior catalogue quality in hot regions or is a real physical areas. A closer examination reveals that the proportion of singles
property. exhibits abrupt spatial fluctuations in some areas (e.g. mid-ocean
Table 3. Testing the hypothesis H0 : Proportion of singles among the clusters is the same in cold and hot regions.
Magnitude Cold regions, H < 0.2 Hot regions, H ≥ 0.2 Fisher test p-value Decision at 0.01 level
No. singles/clusters Prop. singles No. singles/clusters Prop. singles
4.0 ≤ m < 4.5 48 085/49 906 0.963 4216/4422 0.953 0.001 Reject H0
4.5 ≤ m < 5.0 39 654/44 655 0.888 6430/7500 0.857 7 × 10−14 Reject H0
5.0 ≤ m < 5.5 12 238/16 872 0.725 2477/3638 0.681 9 × 10−8 Reject H0
ridges) over hundreds of kilometres. These fluctuations are caused observed values of A (while clusters with larger magnitude gap
by local tectonic and physical settings, such as transition from trans- may artificially become singles). Hence, the reported difference in
form to extension faulting, but are not the focus of this study. The magnitude gap might be influenced to some extent by inferior cata-
relation between the proportion of singles and heat flow is further logue quality in hot areas. However, Zaliapin & Ben-Zion (2013b)
illustrated in Fig. 10(a) that shows pS as a function of strain tensor reported lower magnitude gap in hot regions in a local study in
parameters (S, I2 ). The large-scale averaging used in this analysis southern California, where the quality of catalogues is comparable
clearly demonstrates an increased proportion of singles within cold in both cold and hot regions. We therefore believe that the magni-
areas. The comparison of pS with heat flow and strain rate ten- tude gap difference between hot and cold areas is a real phenomenon

sor parameters using Spearman’s correlation and GLM approach is that will be confirmed in future studies with better catalogue
summarized in Tables 5, 6 and Figs D1(p)–(r). quality.
4.4 Foreshocks and aftershocks 4.5 Structure of earthquake families

The global spatial distribution of the proportion pF of foreshocks Consider a tree T that represents an earthquake family as described
among foreshocks and aftershocks is shown in Fig. 9(b). The pro- in Section 3. The tree consists of a collection of vertices V = {vi },
portion is visibly higher in areas with high heat flow, with typical i = 1,. . . , N each of which represents an earthquake, and edges E
values pF > 0.2, while in areas with low heat flow the typical propor- = {ei }, i = 2, . . . , N such that edge ei connects earthquake i to
tion is very small, pF < 0.1. The increased production of foreshocks its parent that also belongs to the tree T. Here we index the family
in hot regions is confirmed by the analysis of Fig. 10(b) that shows earthquakes in the order of their occurrence time: i < j if and only if
the value of pF averaged for different combinations of strain rate ti < tj . By construction (see Section 3), the parent of the first event
tensor’s style S and second invariant I2 . This result corroborates the in the family does not belong to the same family, and hence the first
earlier regional finding of Zaliapin & Ben-Zion (2013b) that the event does not have an associated edge within T. All other events
number and proportion of foreshocks increases with the heat flow have a single parent from the same family. Hence each tree consists
in southern California, the observation of McGuire et al. (2005) on of N vertices and N−1 edges. We refer to the first event in the family
large proportion of foreshocks within the swarms along the Pacific as the root. Denote by C(i) the number of children of vertex i within
Rise Transform fault, as well as findings of Kagan et al. (2010) and T, and by Np = #{i: C(i) > 0} the number of parental vertices
Chu et al. (2011) based on fitting a branching model to the global within T. We consider two statistics of a time-oriented rooted tree
seismicity. that represent an earthquake family: the average family branching
Spatial patterns similar to those reported in Figs 9(a) and (b) are B and the average leaf depth d (Zaliapin & Ben-Zion 2013b). The
also seen for other examined cluster characteristics. For instance, average family branching B is the average number of offspring per
Fig. A3(d) shows a worldwide map of the aftershock magnitude gap parental vertex of the tree T:
A defined for families with aftershocks as the difference between
the magnitudes of the main shock and the largest aftershock. This 1
B= C(i). (9)
analysis is done for all families with main shock magnitude m ≥ 5. Np i
The gap is generally larger within cold regions, with typical value of
A ≈ 0.8, while in hot regions it is typically smaller, A ≈ 0.55. This The average leaf depth d is the average number of edges between
observation is corroborated in Fig. 10(c) that shows the aftershock a leaf and the tree root. Namely, if di denotes the number of edges
magnitude gap as a function of strain rate tensor parameters (S, I2 ); between vertex i and the tree root, then
the domain of low gap values is similar of that of high heat flow 1
d= di . (10)
shown in Fig. 3(c). The comparison of pF and A with heat flow N − Np i:C(i)=0
and strain rate tensor parameters using Spearman’s correlation and
GLM approach is summarized in Tables 5, 6 and Fig. D1(s)–(x). It is natural to expect the leaf depth and family branching to be
We note that the values of the aftershock magnitude gap reported negatively correlated. An intuitive justification for such reciprocal
here are lower than the value A ≈ 1 suggested by the Båth law (Båth relation comes from the observation that for a tree with constant
1965; Shcherbakov & Turcotte 2004). This deflation is artificial branching and leaf depth, that is with C(i) = C given that C(i) > 0
and is due to the fact that we consider families with main shock and di = D given that C(i) = 0, we have N − Np = CD .
magnitude m ≥ 5, which is only one unit above the magnitude Zaliapin & Ben-Zion (2013b) showed that the values of B and
cut-off mmin = 4 selected for this study. Notably, the difference in d are strongly coupled with the heat flow in southern California.
the magnitude gap A between hot and cold areas is only seen Specifically, the average leaf depth increases while the average
for intermediate-magnitude clusters with m < 6; the difference branching decreases as the heat flow increases. The same general
disappears for large-magnitude clusters with m > 6 (results not trend is observed on the global scale. Fig. 11 shows the values of d
shown). and B averaged for different family sizes N in hot and cold regions.
The magnitude gap is affected by the catalogue completeness It is seen that the average leaf depth d is significantly larger, and the
magnitude, since a higher completeness magnitude leads to smaller family branching B is significantly smaller, in hot areas compared
Figure 9. Global spatial distribution of selected earthquake cluster statistics. (a) Proportion pS of singles among regular clusters. (b) Proportion pF of foreshocks
among foreshocks and aftershocks. (c) Average leaf depth corrected for cluster size, dN , for families with size 5 ≤ N ≤ 20.
to cold areas. We also notice that (i) the difference between hot and Next we focus on the spatial distribution of the average leaf depth
cold regions (difference between red and blue lines) is increasing d and family branching B. The values of both statistics depend on
with the family size and (ii) the values of both statistics increase the family size (Fig. 11), which can contaminate spatial analysis
with the family size. as the family size N significantly varies from region to region, as
Figure 10. Average values of selected cluster statistics as a function of strain rate tensor’s style S and second invariant I2 . (a) Proportion pS of singles among
clusters. (b) Proportion pF of foreshocks among foreshocks and aftershocks. (c) Aftershock magnitude gap A . (d) Average leaf depth corrected for cluster
size, dN , for families with size 5 ≤ N ≤ 20.
(a)
0.8
High heat flow, H > 0.2
Average leaf depth, log10 <d>

0.7 Low heat flow, H < 0.2
0.6
0.5
0.4
0.3
0.2
0.1

3 4 5 6 7 8 9 [10, 20) [20, 50) >50
Family size, N
(b)
0.8
High heat flow, H > 0.2
0.7 Low heat flow, H < 0.2
Family branching, log10 B
0.6
0.5
0.4
0.3
0.2
0.1
3 4 5 6 7 8 9 [10, 20) [20, 50) >50
Family size, N
Figure 11. Average leaf depth d (a) and family branching B (b) as a function of family size N for regions with high (H > 0.2, red solid line and circles) and
low (H < 0.2, blue dashed line and squares) values of the heat flow H. Notice positive trend in both examined characteristics with family size N.
documented in Fig. 8. A least-square regression analysis suggests correlation and GLM approach is summarized in Tables 5, 6 and
that the examined statistics have the following relation to the family Figs D1(y)–(ad). The results confirm that heat flow exerts the pri-
size N in the intermediate size range 5 ≤ N ≤ 20: mary control on the values of these two statistics.
log10 d = 0.35 log10 N + d N , log10 B = 0.5log10 N + B N , (11)

5 DISCUSSION
where dN and BN are respective (non-Gaussian) regression errors
with zero mean. Fig. 12 confirms that the average values of dN and Clarifying whether earthquake dynamics follows universal laws or
BN are fairly independent of the family size in the range 5 ≤ N exhibits different forms related to physical properties of the litho-
≤ 20. By regression construction, the error variables dN and BN sphere is among the main problems of statistical seismology. This
describe variability of the initial statistics d and B, respectively, not study supports earlier results mentioned in the introduction on the
explained by the family size N. Fig. 9(c) show the global spatial existence of non-universal region-specific behaviour of seismic-
distribution of dN for families with sizes 5 ≤ N ≤ 20. This analysis ity. This is done by extending the analysis of Zaliapin & Ben-Zion
confirms our earlier observation: despite some geographic fluctu- (2013a,b) of earthquake clusters in southern California to the global
ations cold regions have a typical value log10 dN ≈ −0.05 that is scale using data from the NCEDC worldwide catalogue for the
consistently smaller than a typical value of the hot regions, log10 dN period 1975–2015. One general difficulty in demonstrating robust
≈ 0.15. The spatial distribution of BN (Fig. A3e) has an inverted differences in properties of earthquakes in different regions is varied
pattern: despite some geographic fluctuations cold regions have a quality of seismic catalogues in different areas. This problem may
typical value log10 BN ≈ 0 that is consistently larger than a typi- be overcome by applying techniques and parameters not sensitive
cal value of hot regions, log10 BN ≈ −0.15. Figs 10(d) and A4(b) to variable location errors and completeness magnitudes (Zaliapin
show the average values of dN and BN , respectively, as functions of & Ben-Zion 2015). We return to this issue below.
strain rate tensor parameters S and I2 . The comparison of dN and BN We use the nearest-neighbour approach (Baiesi & Paczuski 2004;
with heat flow and strain rate tensor parameters using Spearman’s Zaliapin et al. 2008; Zaliapin & Ben-Zion 2013a) to partition the

Figure 12. Average leaf depth dN corrected for cluster size (panel a) and family branching BN corrected for cluster size (panel b) as a function of family size
N for regions with high (H > 0.2, red solid line and circles) and low (H < 0.2, blue dashed line and squares) values of the heat flow H. Notice the absence of
N-dependent trend in both examined characteristics with family size for the intermediate range 5 ≤ N ≤ 20.
earthquakes reported in the NCEDC global catalogue into individ- shocks (Fig. 8); (9) proportion pS of singles among regular families
ual clusters. We then compare the worldwide space distribution of (Figs 9a, 10a and D1p); (10) proportion pF of foreshocks among
various cluster statistics with global heat flow production (Bird et foreshocks and aftershocks (Figs 9b, 10b and D1s); (11) average leaf
al. 2008) and style of lithospheric deformation indicated by an es- depth corrected for the family size, dN (Figs 9c, 10d, 11a, 12a and
timated strain rate tensor (Kreemer et al. 2014). Our comparison is D1y); (12) average family branching corrected for the family size,
based on (i) spatial maps of selected characteristics in seismically BN (Figs A3e, A4b, 11b, 12b and D1ab) and (13) aftershock mag-
active regions (Figs 1, 2, 6 and 9), (ii) averaged values of the ex- nitude gap A (Figs A3d, 10c, D1y). The results are summarized in
amined characteristics as a function of the strain rate tensor style S Tables 4–6.
and second invariant I2 (Figs 3, 7 and 10), (iii) Spearman rank cor- The cluster structure and statistics estimated by our technique
relation analysis (Table 5, Appendix C) and (iv) Generalized Linear are subject to artefacts related to catalogue uncertainties (Zaliapin
Model analysis (Table 6, Fig. D1). & Ben-Zion 2015). We address potential effects of catalogue in-
We demonstrate that multiple statistics of earthquakes and seis- completeness, varying earthquake intensity, and maximal magni-
micity clusters have spatially dependent distribution, tightly corre- tude on each of the examined statistics and design the analysis
lated with the global heat flow production: (1) earthquake intensity to minimize the possible artefacts. Notably, some of our obser-
(Figs 1a, 3b and D1a); (2) average nearest-neighbour earthquake vations (e.g. increased size of small clusters in Fig. 8c and de-
distance ηbg within the background mode (Figs 4, 6a, 7a and D1d), creased proportion of singles in Figs 9a and 10a) demonstrate a
(3) proportion pB of background events (Figs 4, 6b, 7b and D1g), trend that goes against possible artefacts of catalogue uncertain-
(4) quality Q of background/cluster mode separation (Figs 4, A3c, ties. Furthermore, the results of this study are consistent with those
7c and D1j); (5) threshold η0 that separates the background and of a local analysis of southern California (Zaliapin & Ben-Zion
cluster modes (Figs 4, A3b and D1m); (6) rate of temporal de- 2013a,b) obtained with a high quality catalogue (median location
cay of offspring events (Figs 4 and 5); (7) intensity of repeaters error of 500 m) by Hauksson et al. (2012) and much lower min-
(Fig. 5); (8) Cluster size of intermediate magnitude (m < 6) main imal magnitude of analysis, mmin = 2. The combination of our
Table 4. Summary of examined earthquake and cluster statistics.

Statistic Average valuea in Section Figures
Cold regions Hot regions
(H < 0.2) (H > 0.2)
Earthquake intensity, 1.36 0.33 2.1, 4.1 1a, 3b
Background nearest-neighbour distance, log10 (ηbg ) −4.44 −3.91 4.2 4, 6a, 7a
Proportion of background events, pbg 0.57 0.67 4.2 4, 6b, 7b
Quality of mode separation, Q 0.93 0.96 4.2 4, A3c, 7c
Threshold between background and cluster modes, log10(η0 ) −5.40 −5.07 4.2 4, A3b
Rate of temporal decay of offspring low high 4.1 4, 5
Intensity of repeaters low high 4.1 5
Cluster size, N (for main shocks m < 6) 1.28 1.44 4.3 8
Proportion of singles, pS 0.87 0.82 4.3 9a, 10a
Proportion of foreshocks, pF 0.19 0.33 4.4 9b, 10b
Aftershock magnitude gap, A 0.79 0.67 4.4 10c, A3d
Size-corrected leaf depth, dN −0.02 0.07 4.5 9c, 10d, 11a, 12a
−0.07 −0.16

Size-corrected branching, BN 4.5 11b, 12b, A3e, A3b
a The average values are given here to illustrate the trend of changes between cold and hot zones. All reported differences are highly
significant according to ANOVA test (not shown). The average differences reported here are typically lower than those observed for
individual hot/cold regions in the worldwide maps and maps in the strain rate tensor coordinates (S, I2 ).
Table 5. Spearman’s correlation between the examined cluster statistics and heat flow, deformation type.
Variable Heat flow, H Tensor style, S Strain rate second invariant, I2
Earthquake intensity, −0.38a −0.40 0.12
Background nearest-neighbour distance, log10 (ηbg ) 0.46 0.43 0.001 (0.88)
Proportion of background events, pbg 0.23 0.07 0.16
Quality of mode separation, Q 0.34 0.27 0.09
Threshold between background and cluster modes, log10(η0 ) 0.25 0.30 −0.10
Proportion of singles, pS −0.29 −0.21 −0.04 (10−7 )
Proportion of foreshocks, pF 0.38 0.19 0.37
Aftershock magnitude gap, A −0.12 −0.09 −0.03 (10−4 )
Size-corrected leaf depth, dN 0.33 0.22 0.13
Size-corrected branching, BN −0.30 −0.19 −0.09
a The P-values are indicated in parentheses; no P-value is indicated in case P < 10−10 . See Appendix C for definitions.
Table 6. Generalized Linear Model Analysisa : Coefficient of determination for a GLM E[Y] = β 0 + β 1 X + β 2 X 2 + β 3 I{X > mean(X)} .
Predictor, X
Response, Y Heat flow, H Tensor style, S Strain rate second invariant, I2
Earthquake intensity, 0.20 0.16 0.12
Background nearest-neighbour distance, log10 (ηbg ) 0.29 0.19 0.05
Proportion of background events, pbg 0.09 0.09 0.06
Quality of mode separation, Q 0.18 0.09 0.04
Threshold between background and cluster modes, log10(η0 ) 0.09 0.08 0.02
Proportion of singles, pS 0.04 0.03 0.06
Proportion of foreshocks, pF 0.20 0.08 0.11
Aftershock magnitude gap, A 0.04 0.02 0.01
Size-corrected leaf depth, dN 0.14 0.08 0.05
Size-corrected branching, BN 0.10 0.06 0.03
a See Appendix D for definitions and further detail.
global results and those obtained in the detailed regional study of magnitude occurring in a small number of generations and decay-
Zaliapin & Ben-Zion (2013a,b) indicate clear dependency of seis- ing until merging with the background seismicity. Heterogeneity
mic clustering on the heat flow in the region. The results are con- of stress/strength field (which might create fracture barriers) and
sistent with those obtained by Yang & Ben-Zion (2009) and Enescu generally larger failure threshold in cold regions reduce the trig-
et al. (2009) by different statistical analyses, and with theoretical gering potential, which particularly affects small-to-intermediate
expectations based on a viscoelastic damage rheology model (Ben- magnitude events and results in lower overall offspring production,
Zion & Lyakhovsky 2006). smaller cluster size, and increased proportion of singles. At the
The overall picture emerging from these studies indicate that same time, large events (m > 6) have sufficient potential not only for
there exist two primary types of earthquake clustering. (i) Brittle overcoming the failure threshold but also for significantly disturb-
fracture in cold regions results in burst-like clusters characterized ing the neighbouring stress/strength field and generating long slow
by a prominently large main shock that happens in the very be- decaying aftershock sequences. (ii) Brittle–ductile failure mecha-
ginning of the sequence and triggers multiple offspring of smaller nisms in hot regions result in swarm-like clusters that lack a single
prominent main shock. Instead, they gradually develop, event-by- the global strain rate model data. The paper benefitted from com-
event, by triggering earthquakes of comparable magnitudes. The ments by two anonymous reviewers and Editor Eiichi Fukuyama.
offspring span generally multiple generations in such clusters, but Some plots were made using the Generic Mapping Tools version
decay overall much faster creating a notable temporal gap between 4.5.8 (www.soest.hawaii.edu/gmt; Wessel & Smith 1991). This re-
offspring activity in a fading cluster and future background events. search was supported by the Southern California Earthquake Center
The stress/strength field in hot areas is more homogeneous, and the (Contribution No. 6486). Southern California Earthquake Cen-
failure threshold is generally lower than in cold regions, which fa- ter is funded by National Science Foundation Cooperative Agree-
cilitate triggering potential and allows small-to-medium magnitude ment EAR-1033462 & United States Geological Survey Coopera-
events to have offspring. This leads to increased size of clusters tive Agreement G12AC20038.
(for small-to-intermediate main shocks) and decreased proportion
of singles.
Our findings on preferential occurrence of swarm-like clusters in
hot regions, (prominently including the mid-ocean ridges transform REFERENCES
areas) and their general statistical properties, are consistent with
previous large-scale analyses of oceanic swarms (e.g. McGuire et Agresti, A., 2007. An Introduction to Categorical Data Analysis, Wiley,
400 pp.
al. 2005; Roland & McGuire 2009). Furthermore, our results on

Ampuero, J.P. & Ben-Zion, Y., 2008. Cracks, pulses and macroscopic
seismic intensity, cluster size distribution, background proportions,
asymmetry of dynamic rupture on a bimaterial interface with velocity-
and temporal decay of offspring parallel the findings of Kagan et al. weakening friction, Geophys. J. Int., 173(2), 674–692.
(2010) and Chu et al. (2011) on variations of Epidemic-Type After- Baiesi, M. & Paczuski, M., 2004. Scale-free networks of earthquakes and
shock Sequence (ETAS) model parameters across different tectonic aftershocks, Phys. Rev. E, 69, 066106, doi:10.1103/PhysRevE.69.066106.
zones. Their tectonic zones 1 (active continents) and 4 (trenches) Bailey, I.W., Becker, T.W. & Ben-Zion, Y., 2009. Patterns of co-seismic
roughly correspond to the cold areas of our study; while zones strain computed from southern California focal mechanisms, Geophys. J.
2 (slow-spreading ridges) and 3 (fast-spreading ridges) generally Int., 177(3), 1015–1036.
correspond to the hot areas. Bailey, I.W., Ben-Zion, Y., Becker, T.W. & Holschneider, M., 2010. Quanti-
We propose the effective viscosity of the lithosphere to be the fying focal mechanism heterogeneity for fault zones in central and south-
ern California, Geophys. J. Int., 183(1), 433–450.
main control of the style of earthquake clustering. This is consis-
Bak, P. & Tang, C., 1989. Earthquakes as a self-organized critical phe-
tent with interpretations that swarms reflect migration of fluids or
nomenon, J. geophys. Res., 94, 15 635–15 637.
creep (e.g. Hill 1977; Hainzl 2004; Hainzl & Ogata 2005; Vidale Båth, M., 1965. Lateral inhomogeneities of the upper mantle, Tectono-
& Shearer 2006; Chen et al. 2012), as increasing heat flow and physics, 2, 483–514.
fluid content will reduce the effective viscosity and lead to brittle– Ben-Zion, Y., 2001. Dynamic ruptures in recent models of earthquake faults,
ductile deformation that may include creep. However, the expla- J. Mech. Phys. Solids, 49(9), 2209–2244.
nations based on fluid flow and slow slip events appeal to specific Ben-Zion, Y., 2008. Collective behavior of earthquakes and faults:
detailed micromechanisms involving in general many parameters continuum-discrete transitions, progressive evolutionary changes,
(and expected to have additional consequences beyond swarm gen- and different dynamic regimes, Rev. Geophys., 46(4), RG4006,
eration). In contrast, the simpler term effective viscosity involves a doi:10.1029/2008RG000260.
Ben-Zion, Y. & Lyakhovsky, V., 2006. Analysis of aftershocks in a litho-
continuum-based macroscopic description of the behaviour in a re-
spheric model with seismogenic zone governed by damage rheology,
gion (Ben-Zion & Lyakhovsky 2006), not committing to any micro
Geophys. J. Int., 165(1), 197–210.
mechanism. Bird, P., Liu, Z. & Rucker, W.K., 2008. Stresses that drive the plates from
There is no sharp transition between the clusters of the two pri- below: definitions, computational path, model optimization, and error
mary types, and global seismicity exhibits a wide variety of cluster- analysis, J. geophys. Res., 113, B11406, doi:10.1029/2007JB005460.
ing forms. Nevertheless, the cold and hot environments are clearly Chen, X., Shearer, P.M. & Abercrombie, R.E., 2012. Spatial migra-
distinguishable by the average values and distributions of multiple tion of earthquakes within seismic clusters in Southern California:
cluster characteristics. Our analysis also suggests that the type and evidence for fluid diffusion, J. geophys. Res., 117(B4), B04301,
intensity of lithospheric transformation, as measured by the strain doi:10.1029/2011JB008973.
rate tensor, play a secondary role in determining the earthquake Chu, A., Schoenberg, F.P., Bird, P., Jackson, D.D. & Kagan, Y.Y., 2011.
Comparison of ETAS parameter estimates across different global tectonic
cluster style (see, in particular, Tables 5, 6 and Fig. D1). Exam-
zones, Bull. seism. Soc. Am., 101(5), 2323–2339.
ining multiple complementary statistics not sensitive to artefacts
Corral, A., 2004. Long-term clustering, scaling, and universality in the
produced by common catalogue deficiencies (Zaliapin & Ben-Zion temporal occurrence of earthquakes, Phys. Rev. Lett., 92, 108501,
2015) allows us to have confidence that our main findings will re- doi:10.1103/PhysRevLett.92.108501.
main valid in future analyses with improved catalogue quality and Davidsen, J., Gu, C. & Baiesi, M., 2015. Generalized Omori–Utsu law for
alternative cluster identification techniques (e.g. Roland & McGuire aftershock sequences in southern California, Geophys. J. Int., 201(2),
2009; Zhang & Shearer 2016). Additional analyses of seismicity ac- 965–978.
counting for non-universal space-dependent properties, combined Dempster, A.P., Laird, N.M. & Rubin, D.B., 1977. Maximum likelihood from
with geodetic data on aseismic deformation and modelling, can im- incomplete data via the EM algorithm, J. R. Stat. Soc. B (Methodological),
prove further the understanding of earthquake dynamics and provide 1–38.
Enescu, B., Hainzl, S. & Ben-Zion, Y., 2009. Correlations of seismicity
refined information for seismic hazard assessments.
patterns in Southern California with surface heat flow data, Bull. seism.
Soc. Am., 99(6), 3114–3123.
Gu, C., Schumann, A.Y., Baiesi, M. & Davidsen, J., 2013. Triggering cas-
AC K N OW L E D G E M E N T S cades and statistical properties of aftershocks, J. geophys. Res., 118(8),
4278–4295.
We are grateful to Peter Bird for making the heat flow data pub- Gutenberg, B. & Richter, C.F., 1944. Frequency of earthquakes in California,
licly available. We also thank Corne Kreemer for sharing with us Bull. seism. Soc. Am., 34(4), 185–188.
Hainzl, S., 2004. Seismicity patterns of earthquake swarms due to fluid tics applied to the Coso Geothermal field, Geophys. Res. Lett., 42,
intrusion and stress triggering, Geophys. J. Int., 159(3), 1090–1096. doi:10.1002/2015GL064772.
Hainzl, S. & Ogata, Y., 2005. Detecting fluid signals in seismicity data Shcherbakov, R. & Turcotte, D.L., 2004. A modified form of Bath’s law,
through statistical earthquake modeling, J. geophys. Res., 110(B5), Bull. seism. Soc. Am., 94, 1968–1975.
B05S07, doi:10.1029/2004JB003247. Sornette, A. & Sornette, D., 1989. Self-organized criticality and earthquakes,
Hauksson, E., Yang, W. & Shearer, P.M., 2012. Waveform relocated earth- EPL (Europhys. Lett.), 9(3), 197.
quake catalog for Southern California (1981 to June 2011), Bull. seism. Turcotte, D.L. & Malamud, B.D., 2004. Landslides, forest fires, and earth-
Soc. Am., 102(5), 2239–2244. quakes: examples of self-organized critical behavior, Physica A: Stat.
Hicks, A., 2011. Clustering in multidimensional spaces with applications to Mech. Appl., 340(4), 580–589.
statistical analysis of earthquake clustering, in MSc Thesis, Department Utsu, T., 1970. Aftershocks and earthquake statistics (II)-Further investi-
of Mathematics and Statistics, University of Nevada, Reno, August, 2011. gation of aftershocks and other earthquake sequences based on a new
Hill, D.P., 1977. A model for earthquake swarms, J. geophys. Res., 82(8), classification of earthquake sequences, J. Fac. Sci. Hokkaido Univ., Ser.
1347–1352. VII, 3, 197–266.
Kagan, Y.Y., 1999. Universality of the seismic moment-frequency relation, Utsu, T., Ogata, Y. & Matsuura, R., 1995. The centenary of the Omori
in Seismicity Patterns, their Statistical Significance and Physical Mean- formula for a decay law of aftershock activity, J. Phys. Earth, 43(1),
ing, pp. 537–573, eds Wyss, M., Shimazaki, K. & Ito, A., Birkhäuser 1–33.
Basel. Vidale, J.E., Boyle, K.L. & Shearer, P.M., 2006. Crustal earthquake bursts in

Kagan, Y.Y., Bird, P. & Jackson, D.D., 2010. Earthquake patterns in California and Japan: their patterns and relation to volcanoes, Geophys.
diverse tectonic zones of the globe, Pure appl. Geophys., 167(6–7), Res. Lett., 33, L20313, doi:10.1029/2006GL027723.
721–741. Vidale, J.E. & Shearer, P.M., 2006. A survey of 71 earthquake bursts
Keilis-Borok, V.I., 1990. The lithosphere of the Earth as a nonlinear sys- across southern California: exploring the role of pore fluid pressure fluc-
tem with implications for earthquake prediction, Rev. Geophys., 28(1), tuations and aseismic slip as drivers, J. geophys. Res., 111, B05312,
19–34. doi:10.1029/2005JB004034.
Kendall, M.G. & Stuart, A., 1973. The Advanced Theory of Statistics, Volume Weertman, J., 1980. Unstable slippage across a fault that separates elastic
2: Inference and Relationship, Griffin, ISBN 0-85264-215-6, Sections media of different elastic constants, J. geophys. Res., 85(B3), 1455–1461.
31.19, 31.21. Wessel, P. & Smith, W.H., 1991. Free software helps map and display data,
Kreemer, C., Blewitt, G. & Klein, E.C., 2014. A geodetic plate motion and EOS, Trans. Am. geophys. Un., 72(441), 445–446.
global strain rate model, Geochem., Geophys., Geosyst., 15(10), 3849– Yang, W. & Ben-Zion, Y., 2009. Observational analysis of correlations be-
3889. tween aftershock productivities and regional conditions in the context of
McGuire, J.J., Boettcher, M.S. & Jordan, T.H., 2005. Foreshock sequences a damage rheology model, Geophys. J. Int., 177(2), 481–490.
and short-term earthquake predictability on East Pacific Rise transform Zaliapin, I. & Ben-Zion, Y., 2011. Asymmetric distribution of aftershocks
faults, Nature, 434(7032), 457–461. on large faults in California, Geophys. J. Int., 185(3), 1288–1304.
NCEDC, 2015. Northern California Earthquake Data Center, UC Berkeley Zaliapin, I. & Ben-Zion, Y., 2013a. Earthquake clusters in southern Cal-
Seismological Laboratory, Dataset, doi:10.7932/NCEDC. ifornia I: identification and stability, J. geophys. Res., 118(6), 2847–
Omori, F., 1894. On after-shocks of earthquakes, J. Coll. Sci. Imp. Univ. 2864.
Tokyo, 7, 111–200. Zaliapin, I. & Ben-Zion, Y., 2013b. Earthquake clusters in southern Cal-
Reverso, T., Marsan, D. & Helmstetter, A., 2015. Detection and char- ifornia II: classification and relation to physical properties of the crust,
acterization of transient forcing episodes affecting earthquake ac- J. geophys. Res., 118(6), 2865–2877.
tivity in the Aleutian Arc system, Earth planet. Sci. Lett., 412, Zaliapin, I. & Ben-Zion, Y., 2015. Artifacts of earthquake location errors and
25–34. short-term incompleteness on seismicity clusters in southern California,
Roland, E. & McGuire, J.J., 2009. Earthquake swarms on transform faults, Geophys. J. Int., 202 1949–1968.
Geophys. J. Int., 178(3), 1677–1690. Zaliapin, I. & Ben-Zion, Y., 2016. Discriminating characteristics of tectonic
Rubin, A.M., 2002. Aftershocks of microearthquakes as probes and human-induced seismicity, Bull. seism. Soc. Am., in press.
of the mechanics of rupture, J. geophys. Res., 107(B7), Zaliapin, I., Gabrielov, A., Keilis-Borok, V. & Wong, H., 2008. Clustering
doi:10.1029/2001JB000496. analysis of seismicity and aftershock identification, Phys. Rev. Lett., 101,
Rubin, A.M. & Gillard, D., 2000. Aftershock asymmetry/rupture directivity 018501, doi:10.1103/PhysRevLett.101.018501.
among central San Andreas fault microearthquakes, J. geophys. Res., Zhang, Q. & Shearer, P.M., 2016. A new method to identify earthquake
105(B8), 19 095–19 109. swarms applied to seismicity near the San Jacinto Fault, California, Geo-
Rundle, J.B., Turcotte, D.L., Shcherbakov, R., Klein, W. & Sammis, phys. J. Int., 205(2), 995–1005.
C., 2003. Statistical physics approach to understanding the multiscale
dynamics of earthquake fault systems, Rev. Geophys., 41(4), 1019,
doi:10.1029/2003RG000135. APPENDIX A: SELECTED
Schoenball, M., Davatzes, N.C. & Glen, J.M.G., 2015. Differentiat- CHARACTERISTICS OF LITHOSPHERE
ing induced and natural seismicity using space-time-magnitude statis- AND SEISMICITY

Figure A1. Global heat flow distribution, after (Bird et al. 2008). The map is clipped at H = 0.16 W m−2 , while the maximal reported value if H = 0.3 W m−2 .
Figure A2. Characteristics of the strain rate tensor (Kreemer et al. 2014). (a) Strain rate tensor style S of eq. (3). (b) Strain rate tensor second invariant I2 of
eq. (2).
Figure A3. Worldwide spatial distributions of selected earthquake and cluster statistics. (a) Average hypocentral depth, z. (b) Threshold η0 that separates the
background and cluster modes. (c) Quality Q of separation between the background and cluster modes. (d) Aftershock magnitude gap A . (e) Average family
branching BN corrected for cluster size, for families with size 5 ≤ N ≤ 20.
I. Zaliapin and Y. Ben-Zion
Figure A3 (Continued).
630

Figure A4. Average values of selected cluster statistics as a function of strain rate tensor’s style S and second invariant I2 . (a) Average hypocentral depth, z.
(b) Average family branching BN corrected for cluster size, for families with size 5 ≤ N ≤ 20.
the difference between the magnitudes of the family main shock and
A P P E N D I X B : P R O D U C I N G S PAT I A L the largest aftershock (only defined for families with aftershocks).
MAPS
A spatial map of a selected characteristic is produced via the
In this study we produce spatial maps for selected characteristics of
following steps:
the lithosphere, earthquakes, and earthquake clusters. These char-
acteristics can be partitioned into the following four types: (1a) The value of an individual earthquake characteristic is as-
signed to the event epicentre.
(i) Individual earthquake characteristics—magnitude m, (1b) The value of an individual cluster characteristic is assigned
hypocentral depth z, Baiesi-Paczuski distance ηij to the parent, to the epicentre of the family main shock (single is considered to be
foreshock index IF (i) that equals to unity if earthquake i is a a main shock).
foreshock and zero otherwise, and background index IB (i) that (1c) The value of a regional characteristic is estimated at the
equals to unity if earthquake i belongs to background population, epicentre of each catalogue event with magnitude m ≥ 5 within a
and zero otherwise. circle with radius r = 200 km centred at this event.
(ii) Regional characteristics—earthquake intensity , maximal (2) To obtain an averaged (or maximal) value of the selected
observed magnitude mmax , the position ηbg of the background popu- characteristic at point x, the raw estimation from (1) is averaged (or
lation, and quality Q of separation between background and cluster maximized) within circles of radius r = min (r100 , 100 km), where
mode. r100 is the radius of the circle centred at x that contains 100 events
(iii) Individual cluster characteristics—index IS (k) of being a with magnitude m ≥ 5. The points where the circle of radius 100 km
single that equals to unity if cluster k consists of a single event, and contains less than 5 events of magnitude m ≥ 5 are left transparent.
zero otherwise, size-corrected average leaf depth dN (only defined The bandwidth of such adaptive averaging is inversely related to the
for families), size-corrected average family branching BN (only de- seismic intensity, which leads to emphasizing detailed changes of
fined for families), and the aftershock magnitude gap A equal to the examined earthquake characteristics in high-intensity regions
while applying large-scale smoothing in low-intensity regions. The under the null hypothesis of independence has the Student distri-
100 km limit for the radius of the averaging circle produces coloured bution with N − 2 degrees of freedom (Kendall & Stuart 1973).
bands that are often wider than actual seismicity domains, in partic- For large sample sizes this is very close to the standard normal
ular in the transform and divergent environments. This is done for distribution.
visual convenience. The bootstrap and analytical approaches give practically indis-
(3) An isotropic Gaussian filter is applied to a map from (2). This tinguishable results for our sample sizes. We report the significance
last step is only applied to smooth the maps. It does not disturb the levels (see Table 5) according to the analytical approximation of
global patterns of spatial variability. eq. (C1).
APPENDIX D: GENERALIZED LINEAR

APPENDIX C: SPEARMAN M O D E L A N A LY S I S
C O R R E L AT I O N A N A LY S I S
Generalized Linear Models (GLMs) is an extension of linear re-
Spearman rank correlation is designed to detect possible depen- gression framework to non-normal responses that uses the likeli-
dency between non-Gaussian random samples. Recall that the rank hood approach for model fitting. We use GLMs to complement the
r(Xk ) of an observation from a sample {Xi }, i = 1, . . . , N is defined Spearman correlation analysis of Appendix C. Specifically, we see

as the index of this observation in the sample ordered in increasing how the average values of the examined earthquake statistics can
fashion, from the smallest to the largest observation. For exam- be predicted by non-linear functions of (i) heat flow H [W m−2 ], (ii)
ple, the ranks of the observations from the sample {−5, 2, 1} are strain rate tensor style S, and (iii) strain rate tensor second invariant
{1, 3, 2}. In a sample with non-repeating values, the ranks are I2 . For every examined earthquake statistic Y, we fit three GLMs,
natural numbers from 1 to N. one for each predictor X from the above list:
Consider a paired random sample {Xi , Yi }, i = 1, . . . , N. The
Spearman correlation coefficient ρ for the sample is defined as the μ ≡ E [Y ] = β0 + β1 X + β2 X 2 + β3 I{X > mean (X )} . (D1)
Pearson correlation between the ranked sample values: The model’s right hand side combines a quadratic regression in
r (X, Y ) = corr [r (X i ) , r (Yi )] . X and the Heaviside function for the deviation of X from its average.
The goodness of fit is measured by the coefficient of detemination:
Spearman correlation has several properties that make it con-
venient for establishing non-linear relations in non-Gaussian R 2 = 1 − Var [model residuals] /Var [Y] . (D2)
data: 2
Here Var[] denotes the sample variance. The value of R is inter-
(i) The correlation can detect non-linear relations. Namely, preted as the proportion of variance in Y explained by the model. In
ρ(X, Y) = 1 for any monotone increasing deterministic relation Y = particular, R2 = 1 means that the model does a perfect (determinis-
f(X) and ρ(X, Y) = –1 for any monotone decreasing deterministic tic) prediction of Y, while R2 = 0 suggests that the right hand side
relation Y = f(X); of (D1) has no information about Y.
(ii) The correlation is insensitive to monotone transformations of We consider ten earthquake statistics and three predictors, which
data. Namely, ρ(X, Y) = ρ(f(X), g(Y)) for any monotone increasing results in thirty models. The results are summarized in Table 6
functions f and g; and Fig. D1. The solid lines in Fig. D1 represent model fit; the
(iii) ρ(X, Y) is less sensitive to outliers than the conventional amplitude of a line’s jump corresponds to the importance of the
Pearson correlation; last non-linear term I{X > mean (X)} —the difference in the average of
(iv) For Gaussian data, the Spearman correlation is close to Pear- Y for below-than-average and above-then-average values of X. The
son correlation. analysis suggests that the heat flow is the most powerful predictor
among the three examined ones: there exists only one statistic for
A straightforward way of establishing significance of Spearman which the heat flow does not give the highest R2 —the proportion of
correlation is bootstrap: generating multiple independent ranked singles pS .
paired samples and using them to approximate the distribution of ρ We also observe that the quality of forecast is uniformly low—
under the hypothesis of independence. An alternative, approximate, the coefficient of determination never exceeds 0.3. Nevertheless, all
approach suggests that the quantity models with the heat flow as predictor are highly significant. This
confirms the observations in the main part of the paper that heat
N −2
t =ρ (C1) flow is a significant control for the average value of the examined
1 − ρ2 earthquake statistics.
Figure D1. Generalized Linear Model (GLM) analysis of global seismicity. Each panel refers to forecasting a particular characteristic of seismicity using
either heat flow H (first column—a, d, g, j, m, p, s, y, ab), strain rate tensor style S (second column—b, e, h, k, n, q, t, w, z, ac), or logarithm of strain rate
tensor second invariant log10 (I2 ) (third column—c, f, i, l, o, r, u, x, aa, ad). Grey dots—data points; each data point corresponds to a 0.5◦ × 0.5◦ Earth surface
cell. Black lines—GLM forecast. (a, b, c): Seismic intensity [events/year]; the analysis is done for log10 ; (d, e, f): Average distance in background mode,
log10 (ηbg ); (g, h, i): Proportion of background events, pbg ; (j, k, l): Quality of separation between cluster and background modes, Q; (m, n, o): Threshold that
separates the cluster and background modes, log10 (η0 ); (p, q, r): Proportion of singles among clusters, pS ; (s, t, u): Proportion of foreshocks among aftershocks
and foreshocks, pF ; (v, w, x): Difference between magnitude of the main shock and the largest aftershock, A ; (y, z, aa): Average leaf depth, corrected for
family size, dN ; (ab, ac, ad): Average family branching, corrected for family size, BN .
I. Zaliapin and Y. Ben-Zion
Figure D1 (Continued).
634

Zaliapin 2016

Uploaded by

Copyright:

Available Formats

Zaliapin 2016

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Zaliapin 2016

Uploaded by

Copyright:

Available Formats

Geophysical Journal International

Geophys. J. Int. (2016) 207, 608–634 doi: 10.1093/gji/ggw300

A global classification and characterization of earthquake clusters

Ilya Zaliapin1 and Yehuda Ben-Zion2

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

alternative approach discussed by Ben-Zion (2008, Sections 2 and

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

with b = 1 and mc = 4, then p5 = 0.9. Deficiency of events in 3 E A RT H Q UA K E C L U S T E R S

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

(a) Regular analysis (b) Δ−analysis, Δ=2

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Survival function, S(N)

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

4.4 Foreshocks and aftershocks 4.5 Structure of earthquake families

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Average leaf depth, log10 <d>

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

log10 d = 0.35 log10 N + d N , log10 B = 0.5log10 N + B N , (11)

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Table 4. Summary of examined earthquake and cluster statistics.

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

APPENDIX D: GENERALIZED LINEAR

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

Downloaded from https://academic.oup.com/gji/article/207/1/608/2583622 by guest on 20 June 2023

You might also like