JOURNAL OF GEOPHYSICAL RESEARCH, VOL. ???, XXXX, DOI:10.

1029/,

PostScript file created: November 21, 2021

Importance of small earthquakes for stress transfers and

earthquake triggering
Agnès Helmstetter1,3, Yan Y. Kagan2 and David D. Jackson2
1 Institute of Geophysics and Planetary Physics, University of California, Los Angeles, California 90095-1567
arXiv:physics/0407018v2 [physics.geo-ph] 9 Nov 2004

2 Department of Earth and Space Sciences, University of California, Los Angeles, California 90095-1567
3 Now at Lamont-Doherty Earth Observatory, Columbia University

Abstract. We estimate the relative importance of small and large earthquakes for static
stress changes and for earthquake triggering, assuming that earthquakes are triggered
by static stress changes and that earthquakes are located on a fractal network of dimen-
sion D. This model predicts that both the number of events triggered by an earthquake
of magnitude m and the stress change induced by this earthquake at the location of other
earthquakes increase with m as ∼ 10Dm/2 . The stronger the spatial clustering, the larger
the influence of small earthquakes on stress changes at the location of a future event as
well as earthquake triggering. If earthquake magnitudes follow the Gutenberg-Richter law
with b > D/2, small earthquakes collectively dominate stress transfer and earthquake
triggering, because their greater frequency overcomes their smaller individual triggering
potential. Using a Southern California catalog, we observe that the rate of seismicity trig-
gered by an earthquake of magnitude m increases with m as 10αm , where α = 1.05 ±
0.05. We also find that the magnitude distribution of triggered earthquakes is indepen-
dent of the triggering earthquake’s magnitude m. When α ≈ b, small earthquakes are
roughly as important to earthquake triggering as larger ones. We evaluate the fractal cor-
relation dimension D of hypocenters using two relocated catalogs for Southern Califor-
nia. The value of D measured for distances 0.1 < r < 5 km is D = 1.54 for the Shearer
et al. [2003] catalog and D = 1.73 for the Hauksson et al. [2003] catalog. The value of
D reflects both the structure of the fault network and the nature of earthquake inter-
actions. By considering only those earthquake pairs with inter-event times larger than
1000 days, we can largely remove the effects of short-term clustering. Then D ≈ 2, close
to the value D = 2α = 2.1 predicted by assuming that earthquake triggering is due
to static stress. The value D ≈ 2b implies that small earthquakes are as important as
larger ones for stress transfers between earthquakes, and that considering stress changes
induced by small earthquakes should improve models of earthquake interactions.

1. Introduction of “small” earthquakes. Most researchers include only the

mainshock and its largest foreshock or aftershock to pre-
Large shallow earthquakes are followed by increased seis- dict the location of future aftershocks. The success of this
mic activity known as “aftershocks”. Aftershock sequences approach is significant but limited. Only about 60% of after-
of small earthquakes are less obvious because the aftershock shocks are located where the stress increased after a main-
productivity is weaker, but can be observed after stack- shock [Parsons, 2002]; stress shadows (proposed decrease of
ing many sequences [Helmstetter, 2003]. Several mecha- the seismicity rate where Coulomb stress change is negative)
nisms have been proposed to explain earthquake triggering are seldom or never observed [Marsan, 2003; Felzer et al.,
due to the static stress change induced by a prior event: 2003b]; and the correlation of Coulomb stress change with
rate-and-state friction [Dieterich, 1994], sub-critical crack aftershock locations is rather sensitive to the assumed slip
growth [Das and Scholz, 1981; Shaw, 1993], viscous relax- distribution [Steacy et al., 2004]. Coulomb stress change
ation [Mikumo and Miyatake, 1979], static fatigue [Scholz, from the MW 7.3 Landers earthquake does not explain the
1968], pore fluid flow [Nur and Booker, 1972], or simple sand- triggering of the MW 7.1 Hector-Mine earthquake [Harris
pile SOC models [Hergarten and Neugebauer, 2002]. Ka- and Simpson, 2002] which occurred 7 years after Landers
gan and Knopoff [1987a] proposed that random, Brownian and 20 km away. Several alternative models have been pro-
motion-like stress fluctuations cause Omori-law aftershock posed to explain the triggering of the Hector-Mine earth-
rate decay. Coulomb stress change calculations have been quake: viscoelastic effects [Zeng, 2001], dynamic triggering
used to predict the locations, focal mechanisms and times [Kilb, 2003], or secondary aftershocks [Felzer et al., 2002].
of future earthquakes (see reviews by Harris [1998], Stein Felzer et al. [2002] suggested that Hector-Mine may have
[1999] and King and Cocco [2001]). Because large earth- been triggered indirectly by an aftershock of Landers, i.e.,
quakes modify stress over a much larger area than smaller secondary aftershocks and small earthquakes may be impor-
ones, and because computing Coulomb stress changes re- tant for stress triggering.
quires a good model of slip distribution available only for A few studies have estimated the scaling of the after-
large earthquakes, most studies have neglected the influence shock number with the mainshock magnitude [Solov’ev and
Solov’eva, 1962; Papazachos et al., 1967; Utsu, 1969; Singh
and Suarez, 1988; Yamanaka and Shimazaki, 1990; Davis
and Frohlich, 1991; Molchan and Dmitrieva, 1992; Shaw,
Copyright 2021 by the American Geophysical Union. 1993; Drakatos and Latoussaki, 2001]. In these studies, the
0148-0227/21/$9.00 “mainshock” was defined as the largest event of a sequence,
1
X - 2 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING

and “aftershocks” were then selected in a space-time win- • Earthquakes are triggered by a static stress step in-
dow around the mainshock. These studies proposed that duced by the mainshock. Triggering from dynamic stress
the number of aftershocks increases exponentially ∼ 10αm changes may have a different scaling with magnitude. We
with the mainshock magnitude m, with an exponent α in also assume that the mainshock does not change the stress-
the range [0.65, 1]. Other researchers [Jones et al., 1995;
ing rate, i.e., we neglect the influence of post-seismic viscous
Michael and Jones, 1998; Helmstetter, 2003; Felzer et al.;
2004] later suggested that the same mechanisms may ex- relaxation on the seismicity. We do not impose a relation
plain two kinds of triggering: that of aftershocks by a pre- between the number of aftershocks and the stress change;
vious larger earthquake and that of a large earthquake by a we simply assume that the rate of aftershocks at point ~r
previous smaller one. Indeed, the distribution of times be- at time t after a stress change σ is a function which de-
tween an earthquake (“foreshock”) and a subsequent larger pends only on t and σ. For instance, in the rate-and-state
earthquake follows the Omori law as for usual aftershocks model [Dieterich, 1994], the instantaneous increase in seis-
[Jones and Molnar, 1979], and the magnitude of triggered micity rate is proportional to exp(σ), but the total number
earthquakes is independent of that for the triggering event
of aftershocks (integrating over time) is proportional to the
[Helmstetter, 2003]. In the work of Helmstetter [2003] and
Felzer et al. [2004], “aftershocks” were selected as earth- stress change.
quakes occurring in a space-time window after a “main- • A triggered earthquakes size is independent of the mag-
shock” (any earthquake not preceded by a larger one), and nitude of the triggering event (“mainshock”) as suggested
could be larger than the mainshock. Using different meth- by [Helmstetter, 2003]. This implies that the crust is ev-
ods (see section 3.4), Helmstetter [2003] found α = 0.8 ± 0.1 erywhere close to failure, such that any small earthquake,
and Felzer et al. [2004] found α = 1 for Southern California triggered by a previous small one, can grow into an event
seismicity.
Other studies measured the exponent α using a statisti- much larger than its trigger.
cal model of seismicity [Ogata, 1989; Kagan, 1991b; Ogata, • We consider earthquakes with rupture width smaller
1992; Console et al., 2003; Zhuang et al., 2004], which as- than the thickness of the seismogenic crust, for which the
sumes that any earthquake can trigger other events with a rupture length L is proportional to its width W .
rate which decays in time according to Omori’s law and
which increases with magnitude as ∼ 10αm . Ogata has • The rupture length scales with the seismic moment as
1/d
coined the term“ETAS” for “epidemic type aftershock se- L(M ) ∼ M0 . For earthquakes smaller than the thick-
quence” to describe this class of models. However, these ness of the seismogenic crust and for a constant stress drop,
models are now being used to describe the statistics of all d = 3. For larger earthquakes, several models have been
earthquakes, not just aftershocks. Here we will use the des- proposed, with d = 1, d = 2 or d = 3 (see Kagan [2004b] for
ignation “ETES” for “epidemic type earthquake sequence”. a review).
Optimizing the likelihood of the ETES model gives an esti-
mation of the model parameter α. When applied to long- • Earthquakes are located on a fractal network of dimen-
time catalogs, this method usually gives a smaller value of sion D. The spatial clustering of seismicity is due both to
α, in the range 0.5-0.7 [Console et al., 2003; Zhuang et al., the geometry of the fault network and to earthquake inter-
2004], than would result by simply counting the number actions. Over long time-scales, aftershocks should cover uni-
of aftershocks. When applied to individual aftershock se- formly the active fault network, and thus should share the
quences [Guo and Ogata, 1997], this method gives from one spatial distribution of the non-correlated background seis-
sequence to another a large variation of α, ranging from 0.2 micity. In our model, we define D as the fractal dimension
to 1.9 with a mean value of 0.86. of the long-term time-independent seismicity, not including
Yamanaka and Shimazaki [1990] explained the value α =
1 observed for interplate earthquakes by assuming that after- short-term clustering due to earthquake interactions.
shocks are located on patches of the mainshock fault plane. For a constant stress drop, the Coulomb stress change at
If the stress drop is independent of magnitude, the density point ~r due to a finite dislocation of length L and width
of aftershocks on the fault plane is a constant. Therefore, W ∼ L depends only on the ratio r/L and the direction Φ
the number of aftershocks is proportional to the area of the (set of 3 angles) between the rupture plane and the fault
mainshock rupture plane, i.e., α = 1. Hanks [1992] used the plane at point ~r on which we compute the stress [Kagan,
same model to show that, if α is equal to the exponent b of 1991c]. This means that the average stress change at a dis-
the Gutenberg-Richter (GR) law, small earthquakes are just tance r proportional to L from an earthquake of length L
as important as larger ones in redistributing tectonic forces
on faults. Helmstetter [2003] later generalized this model is independent of this earthquake’s magnitude. The only
by assuming that earthquakes are located on a fractal net- difference between small and large events is that larger ones
work of dimension D. This model predicts that stress trans- increase the stress over a larger area. We can thus write
fers between earthquakes and earthquake triggering scales the stress change (tensor) as a function σ(r/L, Φ). We first
as ∼ 10Dm/2 . Therefore, if earthquakes are very clustered estimate the scaling of stress transfers between quakes as a
in space with D < 2b, small earthquakes will be more impor- function of the magnitude of the event which increases the
tant than larger ones for transferring static stress between stress. The expectation of the norm of the stress tensor in-
events and thus for earthquake triggering. duced by an earthquake of length L, integrating over the
In this work we revisit the work of Helmstetter [2003] location of all earthquakes is given by
by using a different method to define mainshocks and af-
tershocks, and by discussing the model in more detail. We Z Z∞
then test the model by estimating the fractal correlation di-
mension D of earthquake hypocenters, using two relocated σ(L) = E(||σ(r/L, Φ)||) = dΦ ||σ(r/L, Φ)|| pr (r)dr(1)
,
catalogs for Southern California. We compare the observed Φ 0
value of α with the prediction α = D/2 [Helmstetter, 2003].
where Φ are limits for a set of 3 angles characterizing fault
2. Scaling of stress transfers and earthquake orientation (their exact form depend on the parametriza-
triggering with the mainshock magnitude tion method) and pr (r) is the time-independent density of
earthquakes at a distance r from another earthquake.
We estimate the relative importance of small earthquakes For a fractal distribution,
for stress transfers and for earthquake triggering. Our model
is based on the following assumptions: pr (r) ∼ r D−1 , (2)
 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING X - 3

where D is the correlation dimension of earthquake hypocen- same as for stress transfers. For seismic moment
ters. The function pr (r) in (1) describes the long-term spa-
D/3
tial distribution of seismicity, not including clustering due Naft (M0 ) ∼ M0 , (6)
to aftershocks. Introducing (2) in (1) and using the new
variable x = r/L, we have and for magnitude
∞
Z Z Naft (m) ∼ 10αm with α = 0.5D . (7)
σ(L) = LD dΦ σ(x, Φ)pr (x)dx . (3)
Φ 0
The result (7) does not depend on the physical mechanism
of earthquake triggering, i.e., on the specific form of the
The integral in (3) is now independent of the events length; function f (σ, t) in (5), but is valid for any model in which
the only dependence on magnitude is in the factor LD . The aftershock properties (seismicity rate, location, duration of
integral is convergent both in the near field (the stress has a the aftershock sequence) are functions of the static stress
r −1/2 singularity close to the crack tip which is integrable) change induced by the mainshock, and on the time since
and in the far field if D < 3 because σ(x) decays as 1/x3 for the stress change, and have no other dependance on the
1/3
x ≫ 1. Assuming that L ∼ M0 ∼ 100.5m (for a constant mainshock magnitude. In the next section, we measure the
stress drop and for L proportional to W ), we can rewrite the scaling of the aftershock rate with the mainshock magnitude
average stress σ(M0 ) induced by an earthquake of moment for Southern California seismicity. We then estimate the
M0 at the location of another event (distributed according fractal dimension D of the spatial distribution of hypocen-
to pr (r)) as [Kagan, 1994] ters in section 5 and compare the results with the model’s
prediction (7).
D/3
σ(M0 ) ∼ M0 . (4)

If earthquakes are distributed uniformly in a volume (D = 3. Scaling of earthquake triggering with

3), or if we compute the integral of the stress change induced mainshock magnitude: observations for
by a quake on a volume rather than at the location of other California seismicity
events, then σ(M0 ) ∼ M0 . In other words, stress change
integrated on a volume is proportional to seismic moment, We have measured the average rate of triggered earth-
and is therefore dominated by the largest earthquakes in a quakes (“aftershocks”) following a “mainshock” in Southern
catalog if the exponent β of the cumulative moment distri- California to measure the scaling of the number of triggered
bution is β < 1 (i.e., b < 1.5 for magnitudes). In contrast, earthquakes with the triggering magnitude. We first dis-
if we compute the stress changes at the location of other cuss some potential problems in analyzing aftershock prop-
quakes, and if D ≈ 2, then the increase in the stress change erties. We describe the catalog in section 3.2. We then de-
with M0 is balanced by the decreased number of events pro-
−2/3 fine “mainshocks” and “aftershocks” and explain our declus-
portional to M0 with seismic moment (equivalent to GR tering procedure. We present our results on the scaling of
law with b = 1). In this case D = 3β, small earthquakes are earthquake triggering with the magnitude of the triggering
just as important as larger ones for stress transfers between
earthquakes. event in section 3.4, and findings on the magnitude distri-
If earthquake triggering is due to static stress changes, bution of triggered earthquakes in section 3.5
the seismicity rate N (r, Φ, t) at point (r, Φ) and time t can
be written as a function f (σ, t) of the stress change σ and 3.1. Problems with estimating aftershock properties
of the time t after the stress change. The function f (σ, t) Several properties of aftershocks are difficult, if not im-
describes any physical mechanism of earthquake triggering possible, to estimate and may be very sensitive to the pa-
due to static stress changes, such as rate-and state friction
[Dieterich, 1994], sub-critical crack growth [Das and Scholz, rameters of the declustering procedure. It is particularly
1981; Shaw, 1993], viscous relaxation [Mikumo and Miy- difficult to define the total number of aftershocks. If earth-
atake, 1979], static fatigue [Scholz, 1968], or pore fluid flow quake triggering follows Omori’s law ∼ 1/tp , with p ≥ 1 and
[Nur and Booker, 1972]. The cumulative number of earth- without cut-off at short times, there is an infinite number
quakes Naft (L) triggered by an earthquake of length L in of aftershocks shortly after the mainshock when the seismic
the time window [c, T ] is given by the integral over time network is saturated. These aftershocks are not reported
and space of the seismicity rate in the catalogs. For p ≤ 1, there are also infinitely many
aftershocks at very long times (between an arbitrary time T
ZT Z Z∞ following the mainshock up to t = ∞) which are mixed up
Naft (L) = dt dΦ f (σ, t)pr (r)dr with non-correlated, background quakes. Therefore, we miss
most aftershocks if we select earthquakes in seismicity cat-
c Φ 0 alogs in a finite time window after the mainshock. Because
ZT Z Z∞ it is impossible to measure the total number of triggered
= L D
dt dΦ f (σ, t)pr (x)dx . (5) events, an alternative solution is to measure the rate of af-
tershocks in a finite time window after the mainshock, when
c Φ 0 the catalog is complete and before the rate goes back to the
constant background level. If the temporal variation in the
The minimum time c is introduced to avoid the singular-
ity of the seismicity rate at time t = 0 predicted by some number of triggered events is the same for all mainshock
models of stress triggering (e.g., in the case of the rate-and- magnitudes (as expected if the rate of aftershocks depends
state model [Dieterich, 1994] when a singularity occurs in only on the stress and the time since the stress step), then
the stress field). The maximum time T is needed to regular- the scaling of the total number of aftershocks (exponent α
ize the integral at large times for some models which predict in (7)) can be measured from that of the rate of triggered
an Omori law decay with p ≤ 1 (but we can take T = ∞ in events with the mainshock magnitude.
Dieterich’s [1994] model). As for the stress change σ(L) in It is also impossible to distinguish between the “direct”
(3), the number of triggered earthquakes depends on mag- (triggered by the “mainshock” only) and “secondary” after-
nitude only in the factor LD . The scaling of aftershock pro- shocks (triggered by a previous aftershock from the main-
ductivity with the size of the triggering earthquake is the shock). Because an earthquake is probably not triggered
X - 4 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING

by a single prior quake, but rather by the cumulative ef- catalog (see http://quake.geo.berkeley.edu/ftp/pub/doc/-
fect of all previous events (with a weight depending on the cat5/cnss.catalog.5 for more details). The magnitudes for
time, distance and magnitude of this particular quake), it is smaller earthquakes are generally mL , or local magnitudes
impossible to tell which earthquake was triggered by which estimated from short-period body waves. Our use of magni-
other one. Only a stochastic answer can be obtained, as- tudes other than moment magnitude is a potential problem,
suming a particular model of earthquake interactions: for because the scaling relationships and simple static stress
example, that earthquakes obey the Epidemic Type Earth- models we employ are based on the seismic moment, which
quake Sequences (ETES or ETAS) model [ Zhuang et al., we assume is simply related to magnitude. Hutton and Jones
[1993] show mL is a fairly good unbiased surrogate for mo-
2004]. After estimating the parameters by maximizing the ment magnitude in the magnitude range 4.5 to 6, but there
likelihood of the model, Kagan and Knopoff [1976] and later may be a significant bias for quakes smaller than 4.5. At this
Zhuang et al. [2004] have proposed a method to calculate point we have to assume that any bias does not affect the
the probability that an earthquake was triggered by a pre- results much, but in the future this assumption needs to be
vious one, or that it was a background event. However, if tested using moment magnitudes for smaller earthquakes.
the number of secondary aftershocks is proportional to the
number of direct aftershocks, then we can measure the total 3.3. Selection of mainshocks and aftershocks
(observed) rate of triggered seismicity, including direct and
The objective of the declustering procedure is to select
secondary aftershocks, which will have the same dependence as mainshocks only events not preceded by any large earth-
on the mainshock magnitude as does the number of direct quake; then quakes following mainshocks within a specified
aftershocks [Helmstetter and Sornette, 2003]. time and distance are identified as aftershocks. The space
Rather than by fitting a multi-parameter stochastic clus- and time windows are chosen in order to minimize the influ-
tering model, such as ETES model, we have estimated earth- ence of non-correlated earthquakes. There are always arbi-
quake clustering properties here directly from the earth- trary choices in a declustering procedure, so we tested dif-
quake catalog, with some simple selection criteria. Although ferent values for all adjustable parameters. As a test we
the ETES type models usually describe earthquake cluster- also applied this declustering method to synthetic catalogs
ing well, using them to determine parameters like α can to estimate the accuracy and lack of bias in results.
be problematic. First, the estimates so determined are not Because we want to investigate the temporal decay of trig-
very robust, because the model parameters are strongly cor- gered seismicity and the scaling of aftershock productivity
related and poorly resolved. Second, the parameter esti- with the mainshock magnitude, we want a declustering pro-
mates may be biased. Inevitable errors in magnitude, loca- cedure which makes minimal assumptions about the prop-
tion, and temporary incompleteness after larger events [Ka- erties of aftershock triggering in time, space and magnitude,
gan, 2004a] may each bias earthquake statistics by under- without introducing a-priori any scaling of the number of af-
tershocks or their duration with the mainshock magnitude.
estimating clustering at short distances and times. Kagan
We assume that larger earthquakes influence seismicity in
[1991b, see p. 142 and his Eq. 36] argues that the value of a wider area, proportional to the mainshock rupture area
α obtained by maximum likelihood inversion of ETES pa- [Kagan, 2002a].
rameters depends on the method used to account for the Practically, we select as mainshocks any earthquake of
incompleteness of short-term aftershocks in a catalog. Fur- magnitude mM not in the influence zone RF × TF of a
thermore, the ETES models generally assume isotropic clus- previous event of magnitude m ≥ mM − ∆m. We use
tering, whereas earthquakes occur preferentially on spatially generally ∆m = 1. This choice ensures that aftershocks
clustered faults. Neglecting this spatial clustering can lead of the foreshock are negligible compared to aftershocks of
to underestimating α, because the contribution to the to- the “mainshock” because aftershock productivity increases
tal predicted seismicity rate from all small aftershocks bet- rapidly with the mainshock magnitude. We compute dis-
ter predicts the location of future aftershocks than does tances between earthquake epicenters, because depth has a
the isotropic contribution from the mainshock. A smaller larger error than horizontal coordinates. The spatial influ-
α gives more weight to smaller earthquakes and therefore ence zone RF of a foreshock is defined as
better accounts for the following factors: incompleteness of
the catalog after a large earthquake, the heterogeneity of RF (m) = max(DF , NL,F × L(m)) , (8)
the spatial distribution of aftershocks, fluctuations of after-
shock productivity, and magnitude errors (see Helmstetter where the constant DF accounts for a location’s accuracy,
L(m) is the rupture length [WGCEP, 2003] of an earthquake
et al. [2004] for more details).
of magnitude m
3.2. Data L(m) = 0.01 × 100.5m (km) , (9)
We used the catalog of seismicity for California pro-
vided by the Advanced National Seismic System (ANSS) and NL,F is an adjustable factor. Mainshocks are selected
(available at http://quake.geo.berkeley.edu/anss/catalog- with a rectangular area 32.1◦ < latitude < 37.4◦ and
search.html). This catalog merges data from several net- −121.9◦ < longitude < −114.1◦ slightly smaller (by 0.1◦ )
works (for this work, essentially Southern and Northern than the box used to select foreshocks and aftershocks so as
California and Nevada). We used earthquakes larger than to avoid finite size effects.
a magnitude threshold md = 2, within the time win- We then select as aftershocks of a mainshock all quakes
dow 1980/1/1-2004/10/15, and within the rectangular area of magnitude m ≥ md in the influence zone RA × TA of the
mainshock, or of a previous aftershock of the mainshock,
32◦ < latitude < 37.5◦ and −122◦ < longitude < −114◦ . even if they are larger than the mainshock. The size of a
This zone does not include Long Valley volcanic seismicity. cluster can thus increase with time, due to aftershock diffu-
We checked that the catalog is complete for m ≥ 2 in this sion or due to secondary aftershocks, as in Reasenberg [1985].
area and time period (except at short times after a large We use a space window
earthquake). We removed explosions from the catalog.
When available, for large earthquakes, we used moment RA (m) = max(DA , NL,A × L(m)) . (10)
magnitudes provided by the Harvard catalog (available at
http://www.seismology.harvard.edu/CMTsearch.html, see We use parameters DA and NL,A in (10) generally smaller
Ekström et al. [2003] and references therein). Other- than the parameters DF and NL,F used to define main-
wise we used the magnitudes provided by the ANSS shocks, because we do not want to consider as a mainshock
 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING X - 5

any earthquake influenced by a previous one, and we do not Omori law decay with p ≈ 0.9 is observed, the earthquakes
want to include as aftershocks distant quakes mixed with which we define as “aftershocks” are probably causally re-
the background. Thus, our declustering method has 7 ad- lated to (triggered by) the “mainshock” rather than simply
justable parameters: DA , DF , NL,A , NL,F , TA , TF and ∆m. correlated to it.
With our algorithm, a small fraction of earthquakes are con- Based on analyzing the magnitude distribution for several
sidered as aftershocks of several mainshocks. For instance,
aftershock sequences in Southern California, Helmstetter et
Landers aftershocks are also considered as aftershocks of a
m = 2.8 earthquake which occurred 6 hours before Landers. al. [2004] have proposed a relation between the magnitude
Our method is very similar to the one of Reasenberg of completeness mc (t, mM ) as a function of the mainshock
[1985], except that we do not require that aftershocks be magnitude and of the time (in days) since the mainshock
smaller than the mainshock, and we do not include a priori
hypotheses on either the scaling of aftershock productiv- mc (t, mM ) = mM − 4.5 − 0.76 log10 (t)
ity with the mainshock or on the temporal decrease of the and mc (t, mM ) ≥ 2. (12)
seismicity rate with time. By contrast, Reasenberg [1985]
assumes that aftershock productivity increases as 102m/3 Earthquakes smaller than this threshold are generally not
and that the rate of aftershocks follows the Omori law with
p = 1. detected due to overlapping of seismic records and satura-
The main differences between this work and the previous tion of the network. Of course some fluctuations of mc occur
analysis of [Helmstetter, 2003] are that Helmstetter [2003] from one sequence to another, but this relation gives a good
used a constant value for RF = 50 km (independent of the fit to all sequences we analyzed within 0.3 magnitude units.
foreshock magnitude), Helmstetter [2003] also used ∆m = 0, For Landers m = 7.3, expression (12) predicts that the com-
and selected as aftershocks earthquakes in the influence zone pleteness magnitude recovers its usual value md = 2 about
of the mainshock only (not in the influence zone of a previous 10 days after the mainshock.
aftershock). We also use here a different catalog from [Helm- For each value of mM , we fit the rate of aftershocks
stetter, 2003] (with different minimum magnitude, space and by (11) with p = 0.9, in the time interval [tmin (mM ),
time windows). tmax (mM )], to estimate K(mM ). The minimum time
Our definition of triggered events is also very similar to
that of Felzer et al. [2004], but the parameters (minimum tmin (mM , md ) is the time after which the catalog is complete
magnitude, time interval) and the method used to com- for m ≥ md , estimated using (12). Parameter tmax (mM ) is
pute α differ. The only difference in the algorithm of af- either fixed to 10 days or given by the condition λ(t, mM ) >
tershock selection is that Felzer et al. [2004] selected after- λmin , where λmin = 0.01 day−1 is the seismicity rate be-
shocks within the influence zone of either the mainshock or low which that rate does not obey Omori’s law (11). At
of its largest aftershock, while we select aftershocks in the large times, the seismicity rate goes to a constant level, due
influence zone of both the mainshock and all its aftershocks. to the background rate or the influence of non-correlated
This difference should have only a minor effect, as aftershock aftershock sequences (see Figure 1). For small mM , the
diffusion is very weak if there is any at all [Helmstetter et seismicity rate increases at large times because the cluster
al., 2003]. Secondly, if a mainshock triggers a larger event, size increases with time, as we enlarge the cluster with new
most of the following events will occur in the influence zone
of the largest earthquake, which generally includes the in- earthquakes related to previous ones. The peak at t ≈ 7
fluence zone of the previous “mainshock”. years for 6 ≤ mM < 6.5 and 7 ≤ mM < 7.5 is due to the
M = 7.1 1999 Hector-Mine aftershock sequence, which is
3.4. Scaling of earthquake triggering with the included in the clusters of the 1992 Landers M = 7.3 and
magnitude of the triggering events Joshua-Tree M = 6.1 earthquakes.
Figure 1b shows that aftershock productivity increases
We stack all aftershock sequences which have the same exponentially with mainshock magnitude. The value of
mainshock magnitude mM , for each class of mainshock mag- K(mM ) (representing the seismicity rate per day for m ≥
nitude ranging from 2 to 7 with a bin size of 0.5. We use
md at time t = 1 day after a mainshock of magnitude mM )
a kernel method [Izenman, 1991] to estimate the seismicity
rate, by convolving the logarithm of earthquake times with a is given by
gaussian kernel of width h = 0.1 for md = 2 and h = 0.2 for
md = 3. The results for DF = 3 km, DA = 2 km, NL,A = 2, K(mM ) = K0 10αmM 10−bmd , (13)
NL,F = 3, TF = 1 yr, ∆m = 1 and md = 3 are shown in Fig-
ure 1. In choosing this parameter, we have checked that our with K0 = 0.0033 dayp−1 and α = 1.07. We recover almost
“mainshocks” are almost uniformly distributed in time, and the same α value as Felzer et al. [2004], who uses a similar
are therefore not strongly influenced by other earthquakes. method of aftershock selection with parameters ∆m = 0,
For each mainshock magnitude, the aftershock rate de- TF = 30 days, TA = 2 days, NF = NA = 2, DF = DA = 5
cays with the time since the triggering event approximately km and two values of md = 3.5 and md = 4.5. They mea-
as sured α by estimating the scaling of the total number of
K(mM ) aftershocks in a 2-day period with the mainshock magni-
λ(t, mM ) = , (11) tude. This value of α is close to the GR b-value (see Figure
tp
1b). The b-value measured by mean-square linear regres-
with p ≈ 0.9. This expression (11) is only correct in a finite sion of the cumulative distribution for m ≥ 2 in the time
time window [tmin (mM ), tmax (mM )], for which the catalog window 1980-2004 gives b = 0.94. A maximum likelihood
is complete above md and the proportion of non-correlated method gives b = 1.03. This value of b ≈ α implies that all
events is negligible. The Omori law has been reported in earthquakes in a given magnitude range collectively attain
many observations of aftershock sequences (see Utsu et al. the same importance for earthquake triggering, as suggested
[1995] for a review), and can be derived from several phys-
previously by Agnew and Jones [1991], Michael and Jones
ical models of earthquake triggering. For example, Omori
decay (11) with an exponent p ≈ 0.9 can be reproduced by [1998] and Felzer et al. [2002, 2004]. For α = b, the in-
the rate-and-state model of Dieterich [1994] using a non- crease of aftershock productivity with mM compensates for
uniform stress distribution. Due to our rule of aftershocks the decreased number of earthquakes with magnitude. Even
and mainshocks selection, the seismicity rate after the main- if an earthquake of magnitude m = 7 triggers on average 105
shock is much larger than before. We conclude that, when times more aftershocks than a magnitude 2, an earthquake of
X - 6 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING

any magnitude is as likely to be triggered by a m = 7 earth- sequence is at least 1000 days, independent of the main-
quake as by a m = 2, because there are 105 more m = 2 shock magnitude, as predicted by the rate-and-state model
than m = 7 earthquakes in the catalog. of seismicity [Dieterich, 1994]. This figure also confirms that
The roll-off of the seismicity rate at times t < tmin (mM ) Omori’s exponent does not depend on the mainshock mag-
can be explained by catalog incompleteness at shorter times nitude.
[Kagan, 2004a]. To check this, we have corrected the seis- The instantaneous increase of seismicity rate after a main-
micity rate using GR law with b = 1 to estimate the rate shock, compared to the long-term rate, is at least of a factor
of seismicity for m ≥ md from the observed rate of events 105 , maybe larger if we could detect aftershocks at shorter
for m ≥ mc (mM , t). We have removed from the catalog all times. Using the rate-and-state model [Dieterich, 1994], this
events with m < mc (mM , t), measured the rate of seismicity means that the ratio ∆σ/Aσn (where ∆σ is the maximum
λm≥mc (t, mM ) as a function of time and mainshock magni-
Coulomb stress change, A is a parameter of the rate-and-
tude, and then estimated the rate of m ≥ md earthquakes
state friction law, and σn is the normal stress) is at least
by
equal to log(105 ) = 11.5. Assuming Aσn = 0.4 bar [Toda
and Stein, 2003], this gives a maximum stress of 4.6 bar.
λm≥md (t, mM ) = λm≥mc (t, mM )10mc (t,mM )−md . (14)
3.5. Magnitude distribution of triggered earthquakes
The results are shown in Fig. 2: the rate of seismicity now
follows the Omori law for t ≥ 2 × 10−3 days for all main-
shock magnitudes. The cut-off at t ≈ 10−3 days (86 sec) We have analyzed the magnitude distribution of triggered
is probably due to the fact that earthquakes at such small earthquakes for each class of the mainshock magnitude be-
times cannot be distinguished from the mainshock and are tween 2 and 7. As above, we select aftershocks in a time
therefore not reported in the catalog. window [tmin (mM ), tmax (mM )] such that the catalog is com-
We have tested different values of the parameters of af- plete above md = 2 and that the influence of non-correlated
tershock and mainshock selection (see Table 1). The fitting earthquakes is negligible (λ(t, mM ) > 0.1 day−1 ). The re-
interval (tmax and λmin ) was adjusted so that the seismic- sults are shown in Figure 4. We observe that the magni-
ity rate for t < tmax and λ(t) > λmin decays according to
tude distribution of triggered earthquakes follows the GR
Omori’s law with p ≈ 0.9, as expected for triggered earth-
quakes. All tests, for reasonable values of DA smaller than law with b ≈ 1, independent of the mainshock magnitude.
the location accuracy (2 km for A-quality locations), give For 7 ≤ mM < 7.5 (Hector-Mine and Landers), the b-value
α = 1.05 ± 0.05. There are however large variations of K0 as seems larger than 1. However, if we use md = 3 and a
a function of the parameters of aftershock and mainshock se- smaller minimum time tmin , the magnitude distribution is
lection. The value of α exceeds the value α = 0.8±0.1 found close to the GR law with b = 1. These results imply that
by Helmstetter [2003]. This discrepancy is due to differences a small earthquake can trigger a much larger earthquake.
in the selection of aftershocks and mainshocks. Helmstetter It thus validates our hypothesis that the size of a triggered
[2003] used a constant value of RF = 50 km, independent earthquake is not determined by the size of the trigger, but
of the “foreshock” magnitude. This value RF = 50 is too that any small earthquake can grow into a much larger one
small to exclude distant Landers aftershocks. Helmstetter [Kagan, 1991b; Helmstetter, 2003; Felzer et al., 2004]. The
[2003] also used a larger value of tmax = 1 yr, independent magnitude of the triggering earthquake controls only the
of mM , so that a significant fraction of background events number of triggered quakes.
may have contaminated aftershocks of small mainshocks.
We have applied the same method (selection of main- 3.6. Proportion of triggered events in catalogs
shocks and aftershocks, correction for undetected early af-
tershocks, and fit by Omori’s law) to synthetic catalogs gen- We use the scaling of the average number of triggered
erated with the ETES model. In this model, the seismicity earthquakes per mainshock and the hypothesis that the
rate λ(t, ~r, m) is the sum of a constant background term µ magnitude distribution of all quakes follows the GR law
and of aftershocks from all previous earthquakes: to derive the long-term fraction of aftershocks (assuming
h boundaries in time and space for the definition of after-
shocks). The rate of m ≥ md aftershocks triggered, directly
X
λ(t, ~r, m) = pm (m) µ + k10α(m−m0 )
or indirectly, by an earthquake of magnitude m at time t
ti <t
after this quake follows approximately
θcθ
νL(mi )ν
i
× , (15)
(t + c) 1+θ (|~r − r~i | + L(mi ))1+ν K0 10αm 10−bmd
λm (t) = , (16)
tp
0.5m
where L(m) = 0.01 × 10 is the rupture length, pm (m)
is the magnitude distribution, c, θ, k, α, ν are adjustable with p = 0.9 and K0 = 0.0041 dayp−1 . This relation holds
parameters. Our method recovers the α value of the model at least for times t > 0.001 day (86 sec), for all mainshock
with an accuracy ≈ 0.03, but underestimates the total num- magnitudes, after correcting for the incompleteness of the
ber of aftershocks, because distant aftershocks outside the catalog at short times using eq. (12). For m ≥ 7, the
“influence zone” used for aftershock selection are missed (see Omori law decay holds at least up to T ≈ 100 days. If
results in Table 1). we assume that this “aftershock duration” does not depend
In order to study the rate of aftershocks at large times on the mainshock magnitude, the average total number of
after a mainshock, we have selected aftershocks very close m > 2 aftershocks (including secondary aftershocks) trig-
to the mainshock, in order to minimize the influence of non- gered by m ≥ md mainshocks in the time window c < t < T
correlated earthquake. We used parameters DA = 0.5 km, is
NL,F = 2, DF = 3 km, and NL,F = 3, and we did not
include aftershocks in the influence zone of previous after- ZT
shocks (see model #8 in Table 1). We thus under-estimate T 1−p − c1−p
the number of aftershocks for small mainshocks, because the Naft (m) = λm (t)dt = K0 10αm−bmd .(17)
1−p
influence zone of small events is smaller than the location c
accuracy, and because we miss more secondary distant af-
tershocks for small mainshocks. The results are shown in Averaging over all magnitudes of the mainshock we can com-
Figure 3. We observe that the duration of the aftershock pute the average total number Naft of m ≥ md aftershocks
 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING X - 7

(including secondary aftershocks) per mainshock of magni- Marsan [2004] solved this problem by computing the dis-
tude md < m < mmax . Assuming a GR law with b = 1 and tribution of the stress induced by an earthquake at the loca-
with an upper magnitude cut-off at m = mmax , Naft is given tion of another earthquake, and by studying only the tail of
by the distribution (small absolute value of stress) correspond-
m
ing to distances larger than the rupture length of the event.
Zmax Marsan [2004] concluded that small earthquakes are at least
Naft = p( m)Naft (m)dm (18) as important as larger ones for redistributing stress between
md
events. This result is in agreement with our result for scal-
m max
ing the number of triggered earthquakes with the mainshock
T 1−p − c1−p magnitude. But even in the far field, these calculations of
Z
−bm αm
= K0 b ln(10)10 10 dm the Coulomb stress change induced by an earthquake are
1−p
md very inaccurate, because the accuracy of focal mechanisms
◦
(α−b)md  1−p 1−p is on the order of 30 [Kagan, 2002b]. Another problem in
K0 b10 (α−b)(mmax −md ) T
 −c
= 1 − 10 . [Marsan, 2004] is that he considers earthquakes of magni-
b−α 1−p tude m = 3, which have a rupture length L ≈ 300 m, much
smaller than the average accuracy on vertical coordinates of
The special case α = b gives
4 km. These location errors can significantly bias any esti-
T 1−p
−c 1−p mate of the stress change induced by small m < 5 quakes
Naft = K0 b ln(10)(mmax − md ) . (19) [Huc and Main, 2003].
1−p The observation that earthquake triggering scales with
αm
Using the parameters α = 1.05, b = 1., c = 0.001 day, the magnitude m of the trigger as ∼ 10 with α ≈ b sug-
K0 = 0.0041 day p−1
, and md = 2 estimated for Southern gests that small events should not be neglected in studying
California (see Table 1 line 2) and assuming a maximum stress interactions between quakes. However, directly calcu-
magnitude mmax = 8 and an aftershock duration T = 100 lating the stress induced by a small events is impossible due
days, we obtain Naft = 1.11. In other words, a fraction to the inaccuracy of earthquake locations and focal mecha-
Naft /(Naft + 1) = 53% of m ≥ 2 earthquakes in Califor- nisms. Even if we cannot compute the spatial distribution
nia are aftershocks of a m ≥ 2 earthquake. If we assume of the stress induced by a small earthquake, we can esti-
that Omori’s law with p = 0.9 holds up to T = 10000 days mate how such small events trigger others. On average the
(27 yrs), then Naft = 2.06 (67% of earthquakes are after- total number of events triggered by an earthquake (inte-
shocks). The results do not depend on c if p < 1. The grated over space) will be independent of the earthquake
fraction of aftershocks increases if we include mainshocks source. For instance, we can account for the influence of
with m ≤ 2. Assuming that equation (17) and the GR law small earthquakes in the rate-and-state model of Dieterich
still hold down to md = −1 [Abercrombie, 1995], we obtain [1994] by using for such small events a point source model
Naft = 2 for T = 100 days: 67% of earthquakes of any mag- which, as we know, depends only on the focal mechanism.
nitude are triggered by earthquakes of magnitude m ≥ −1.
We can then compute the seismicity rate on each point of
The fraction of triggered events goes to 1 as the minimum
magnitude md goes to −∞ [Sornette and Werner, 2004]. a grid by integrating on each cell that rate estimated us-
Our results are relatively close to the generic model of ing equation (12) of [Dieterich, 1994] (using space-variable
Californian aftershocks Reasenberg and Jones [1989, 1994] stress for a point-source model). If the size of each cell is
(RJ), which assumed equal to a few km, a point source model approximates m ≤ 4
earthquakes well, because the integral of the seismicity rate
10a+b(m−md ) over each cell depends only slightly on the geometry of the
λm (t) = , (20) rupture fault and the variations of fault slip.
(t + c)p
Another alternative is to use empirical laws like the ETES
with c = 0.05 days, p = 1.08, b = 0.91, a = −1.67 [Reasen- model [Kagan and Knopoff, 1987b; Felzer et al., 2003a;
berg and Jones, 1994]. Their model gives K0 = 10a = 0.017 Helmstetter et al., 2004], rather than stress calculations and
day−1 for md = 2, but with smaller α and b values (α is physical models of stress interactions. We can estimate the
assumed to be equal to b), so that the rate of m ≥ 2 after- spatial distribution of future aftershocks for a large earth-
shocks at t = 1 day and for m = 7 is λm (t) = 758 day−1 for quake by smoothing the locations of early aftershocks. A
RJ model and λm (t) = 729 day−1 for our model (line #2 in magnitude m = 7 quake has enough aftershocks in the
Table 1). first hour to estimate the distribution of future aftershocks
[Helmstetter et al., 2004]. In contrast, predictions based on
4. Scaling of stress transfers with the Coulomb stress change calculations require a good model of
slip distribution on the fault, which is not available until a
mainshock magnitude few hours after the mainshock at best.
Computing stress changes induced by a small earthquake
at the location of another quake is very difficult. The stress 5. Spatial distribution of seismicity
field created by an earthquake is sensitive to the slip dis-
tribution on the fault, at least for small distances from the We have estimated the distribution of distances be-
fault plane [Steacy et al., 2004]. Slip distribution is usually tween hypocenters pr (r), using the relocated catalogs for
available only for large m ≥ 6 earthquakes in California. Southern California by Hauksson et al. [2003] (HCS) and
For small earthquakes, we can use a point-source model if
we know the focal mechanism, to calculate stress for dis- Shearer et al. [2003] (SHLK). These catalogs apply waveform
tances from the fault much larger than the rupture dimen- cross-correlation to obtain precise differential times between
sion. We can also estimate the most likely focal plane from nearby events. These times can then be used to greatly im-
the fault orientations in the area or from the orientation of prove the relative location accuracy within clusters of sim-
the tectonic stress. Then we can model the rupture by a ilar events. Locations in these two catalogs do not always
rectangular dislocation with a uniform or tapered slip. But agree in detail, reflecting their different modeling assump-
this simple source model would still be incorrect close to the tions and seismic velocity structures. However, their overall
fault, where most aftershocks are located. agreement is quite good, particularly when compared to the
X - 8 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING

standard catalog locations. In many regions, these new cat- increases, the fraction of small distances will decrease. The
alogs resolve individual faults in what previously appeared fractal dimension of pr (r, t) increases between D ≈ 0 at
to be diffuse earthquake clouds. In both catalogs, we have times t = 5 minutes up to D ≈ 2 for t = 2500 days (see
selected only m ≥ 2 earthquakes relocated with an accuracy Figure 6). This maximum inter-event time of 2500 days is
of ǫh and ǫz smaller than 100m. In the HCS catalog, there long enough so that earthquake interactions are negligible
are 71943 m ≥ 2 earthquakes between 1984 and 2002, out compared to tectonic loading, and a very small fraction of
of which 24127 (34%) are relocated with ǫh < 0.1 km and earthquake pairs with τ > 1000 days belong to the same af-
ǫz < 0.1 km. In the SHLK catalog, there are 82442 m ≥ 2 tershock sequence. This value D = 2, measured for t = 2500
earthquakes in the same period [1984, 2002], out of which days, can thus be interpreted as the fractal dimension of the
33676 (41%) are relocated with ǫh < 0.1 km and ǫz < 0.1 active fault network.
km. The probability density function of distances pr (r) be-
tween hypocenters is close to a power-law pr (r ∼ r D−1 in
the range 0.1 ≤ r ≤ 5 km. The correlation fractal dimen- 6. Conclusion
sion (measured by least-square linear regression of log(r)
and log(pr (r)) for 0.1 ≤ r ≤ 5 km) is D = 1.54 for SHLK Although large earthquakes are much more important
and D = 1.73 for HCS (see black lines in Figure 5). The than smaller ones for energy release, small quakes have col-
faster decay for r < 0.1 km is due to location errors, and lectively the same influence as large ones for stress changes
the roll-off for distances r > 5 km is due to the finite thick- between earthquakes, due to seismic spatial clustering. Be-
ness of the seismogenic crust. The difference in the D-value cause the stress drop is constant, the stress change induced
between the two catalogs may come from larger location er- by an earthquake of magnitude m at the location of other
rors in HCS, but also from errors in estimating the D-value earthquakes (located on a fractal network of dimension D)
(see Kagan [2004c] for a discussion of errors and biases in increases with m as ∼ 10Dm/2 . Measuring directly the scal-
determining the fractal correlation dimension). ing of such stress change with magnitude is difficult, because
To check the accuracy of the D-value, we tested a syn- the accuracy of earthquake locations and focal mechanisms
thetic catalog, generated with the same number of events is limited. We can, however, estimate indirectly the rela-
as the real catalog. We used random longitudes and lat- tive importance of small and large events for stress transfer
itudes, with a uniform epicentral distribution within the by studying the properties of triggered seismicity. If quakes
same boundaries as real data. Depths in the synthetic are triggered by static stress changes, then the number of
catalog were chosen by shuffling the depths of earthquakes such triggered events should also scale as ∼ 10Dm/2 with
from the real catalog to keep the same depth distribution.
the magnitude of the trigger.
The fractal dimension in the range 0.1 ≤ r ≤ 5 km was
We have measured the average seismic rate triggered by
D = 2.93 ± 0.04: a little smaller than the value D = 3 ex-
pected for a purely uniform distribution in a volume. This an earthquake for Southern California seismicity. We found
discrepancy is due to three factors: the finite number of that the rate of triggered events decays with time accord-
events, the non-uniform distribution of depths, and finite ing to Omori’s law ∼ 1/(t + c)p with p = 0.9 and c < 3
size effects. minutes (after correcting for the increase in the complete-
The spatial clustering of earthquake hypocenters mea- ness threshold after a large mainshock, see eq. 12). This
sured for the full catalog is due both to the fault network decay is independent of the mainshock magnitude m for
structure and to earthquake interactions. The latter mod- 2 < m < 7.5. We also found that the magnitude of trig-
ify long-term spatial distribution by increasing the fraction gered aftershocks follows the Gutenberg-Richter law with
of small inter-event distances, that is decreasing the frac- b = 1 and is independent of the mainshock magnitude. The
tal dimension [Kagan, 1991a]. The clustering of aftershocks rate of triggered quakes increases with m as ∼ 10αm , with
around mainshocks, and of secondary aftershocks close to an exponent α = 1.05 ± 0.05.
the direct aftershock of the mainshocks, and so on, can cre- We have measured the fractal correlation dimension D for
ate a fractal distribution. This results from the scaling of the two different catalogs of relocated earthquakes in Southern
aftershock zone with magnitude coupled with the GR law, California, given by the exponent of the cumulative distri-
without any underlying fractal fault structure [Helmstetter bution of the distances between all pairs of hypocenters.
and Sornette, 2002]. This fractal dimension D ≈ 1.6, measured in a limited time-
Over very long time intervals, triggered earthquakes window, characterizes the spatial clustering of earthquakes
should be located uniformly on the fault network, and the due to fault network structure and earthquake interactions.
fractal dimension of the whole catalog should reach a con- By using only earthquake pairs with inter-event times larger
stant value. In our model described in equation (5), pr (r) than a threshold to compute D, we have tried separating the
represents the “long-term” time-independent spatial dis- effects that fault geometry and earthquake triggering have
tribution of seismicity. Distribution pr (r) gives the aver- upon spatial clustering. For inter-event times larger than
age density of earthquakes at a distance r from the main- 1000 days, we obtain D ≈ 2. Thus, our result α ≈ D/2
shock due to fault structure geometry and the faults time- supports the assumption that earthquakes are triggered by
independent heterogeneity. This does not take into account
static stress from past events.
earthquake interactions. The effect of the mainshock, which
The fact that α is nearly equal to the b-value has im-
increases or decreases the seismicity rate at close distances
by modifying the stress, is described in the factor f (σ, t) portant consequences for earthquake triggering. It means
in (5). Therefore, we should not use in (5) the function that equal magnitude bands contribute equally to stress
pr (r) estimated from the catalogs of Southern California at the location of a future hypocenter, e.g., earthquakes
seismicity because these catalogs cover less than 20 years with 2 < m < 5 have collectively the same importance as
and contain a majority of aftershocks. If we had included 5 < m < 8 events for stress changes and for triggering, be-
that function, we would be double-counting earthquake in- cause the frequency of small earthquakes compensates for
teractions. their lower triggering potential. Even if explicit stress cal-
To estimate and remove the time dependence of the spa- culations based on cataloged earthquakes have limited ac-
tial distribution of inter-event distances, we have measured curacy, we can estimate how small earthquakes affect earth-
that distribution pr (r, t) using only earthquake pairs with quake triggering. This can be done with a spatial resolution
an inter-event time τ in the range [t, t + dt]. The results of a few km, using an ETES model or the more physical
are shown in Figure 5. As the minimum inter-event time rate-and-state model.
 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING X - 9

Acknowledgments. Helmstetter, A., and D. Sornette (2003), Importance of direct and

We acknowledge the Advanced National Seismic System, Egill indirect triggered seismicity in the ETAS model of seismicity,
Hauksson, Peter Shearer, and the Harvard group for the earth- Geophys. Res. Lett., 30, 1576, doi:1029/2003GL017670.
quake catalogs used in this study. We thank Emily Brodsky, Jim Helmstetter, A., G. Ouillon and D. Sornette (2003), Are after-
Dieterich, Karen Felzer, Paolo Gasperini, Jean-Robert Grasso, shocks of large Californian earthquakes diffusing?, J. Geophys.
and Max Werner for useful discussions. We are grateful to the Res., 108, 2483, 10.1029/2003JB002503.
associate editor Massimo Cocco, and to the reviewers Bruce Shaw, Helmstetter, A., Y. Y. Kagan and D. D. Jackson (2004), Compar-
Susanna Gross and Ruth Harris for useful suggestions. We ac- ison of short-term and long-term earthquake forecast models
knowledge David Marsan for sending us a preprint of his paper. for southern California, to be submitted to Bull. Seism. Soc.
We are grateful to Kathleen Jackson for editing the manuscript. Am.
This work is partially supported by NSF EAR 0409890 and by the Hergarten, S. and H. J. Neugebauer (2002), Foreshocks and Af-
Southern California Earthquake Center (SCEC). SCEC is funded tershocks in the Olami-Feder-Christensen Model, Phys. Rev.
by NSF Cooperative Agreement EAR-0106924 and USGS Coop- Lett., 88, 238501.
erative Agreement 02HQAG0008. The SCEC contribution num- Izenman, A. J. (1991), Recent developments in nonparametric
ber for this paper is 805. density estimation, J. Am. Stat. Assoc., 86, 205-224.
Huc, M., and I. G. Main (2003), Anomalous stress diffusion in
earthquake triggering: Correlation length, time dependence,
and directionality, J. Geophys. Res., 108, 2324, 1-12.
References Hutton L.K., and L. M. Jones (1993), Local magnitudes and ap-
parent variations in seismicity rates in southern California,
Abercrombie, R. E. (1995), Earthquake source scaling relation- Bull. Seism. Soc. Am., 83, 313,329.
ships from −1 to 5 ML using seismograms recorded at 2.5-km Jones, L. M., and P. Molnar (1979), Some characteristics of fore-
depth, J. Geophys. Res., 100, 24,015-24,036. shocks and their possible relationship to earthquake prediction
Agnew, D. C. and L. M. Jones (1991), Prediction probabilities and premonitory slip on fault, J. Geophys. Res., 84, 3596-3608.
from foreshocks, J. Geophys. Res., 96, 11,959-11971. Jones, L. M., R. Console, F. di Luccio and M. Murru (1995), Are
Console, R., M. Murru and A. M. Lombardi (2003), Refining foreshocks mainshocks whose aftershocks happen to be big?,
earthquake clustering models, J. Geophys. Res., 108, 2468, EOS, 76 (abstract), 388.
doi:10.1029/2002JB002130. Kagan, Y. Y. (1991a), Fractal dimension of brittle fracture, J.
Das, S. and C. H. Scholz (1981), Theory of time-dependent rup- Nonlinear Sc., 1, 1-16.
ture in the Earth, J. Geophys. Res., 86, 6039-51. Kagan, Y. Y. (1991b), Likelihood analysis of earthquake cata-
Davis, S. D. and C. Frohlich (1991), Single link cluster analysis, logues, Geophys. J. Int., 106, 135-148.
synthetic earthquake catalogues, and aftershock identification, Kagan, Y. Y. (1991c), 3-D rotation of double-couple earthquake
Geophys. J. Int., 104, 289-306. sources, Geophys. J. Int., 106, 709-716.
Dieterich, J. (1994), A constitutive law for rate of earthquake pro- Kagan, Y. Y. (1994) Distribution of incremental static stress
duction and its application to earthquake clustering, J. Geo- caused by earthquakes, Nonlinear P. Geophys., 1, 172-181.
phys. Res., 99, 2601-2618. Kagan, Y. Y. (2002a), Aftershock zone scaling, Bull. Seismol.
Drakatos, G. and J. Latoussakis (2001), A catalog of aftershock Soc. Amer., 92, 641-655.
sequence in Greece (1971-1997): Their spatial and temporal Kagan, Y. Y. (2002b), Modern California earthquake catalogs
characteristics, Journal of Seismology, 5, 137-145. and their comparison, Seism. Res. Lett., 73(6), 921-929.
Ekström, G., A. M. Dziewonski, N. N. Maternovskaya and M. Kagan, Y. Y., (2004a) Short-term properties of earthquake cat-
Nettles (2003), Global seismicity of 2001: centroid-moment alogs and models of earthquake source, Bull. Seismol. Soc.
tensor solutions for 961 earthquakes, Phys. Earth Planet. In- Amer., 94(4), 1207-1228.
ter., 136, 165-185. Kagan, Y. Y. (2004b), Earthquake slip distribution, submitted to
Felzer, K.R., T.W. Becker, R.E. Abercrombie, G. Ekström and J. Geophys. Res.
J.R. Rice (2002), Triggering of the 1999 MW 7.1 Hector Mine Kagan, Y. Y. (2004c), Earthquake spatial distribution: correla-
earthquake by aftershocks of the 1992 MW 7.3 Landers earth- tion dimension, in preparation.
quake, J. Geophys. Res., 2190, 10.1029/2001JB000911. Kagan, Y. Y. and L. Knopoff (1976), Statistical search for non-
Felzer, K. R., R. E. Abercrombie, and G. Ekström (2003a), Sec- random features of the seismicity of strong earthquakes, Phys.
ondary aftershocks and their importance for aftershock predic- Earth Planet. Inter., 12, 291-318.
tion, Bull. Seis. Soc. Am., 93, 1433-1448. Kagan, Y. Y., and L. Knopoff (1987a), Random stress and earth-
Felzer, K.R., R.E. Abercrombie and E. E. Brodsky (2003b), Test- quake statistics: Time dependence, Geophys. J. Roy. Astr.
ing the stress shadow hypothesis, Eos Trans. AGU, 84(46), Soc., 88, 723-731.
Fall Meet. Suppl., Abstract S31A-04. Kagan, Y. Y., and L. Knopoff (1987b), Statistical short term
Felzer, K. R., R. E. Abercrombie, and G. Ekström (2004), A com- earthquake prediction, Science, 236, 1563-1467.
mon origin for aftershocks, foreshocks, and multiplets, Bull. Kanamori, H. and D. Anderson (1975), Theoretical basis of some
Seis. Soc. Am. , 94, 88-99. empirical relations in seismology, Bull. Seism. Soc. Am., 65,
Guo, Z. and Y. Ogata (1997), Statistical relations between the 1073-1095.
parameters of aftershocks in time, space and magnitude, J. Kilb, D. (2003), A strong correlation between induced peak dy-
Geophys. Res., 102, 2857-2873. namic Coulomb stress change from the 1992 M 7.3 Landers,
Hanks, T. C. (1992), Small earthquakes, tectonic forces, Science, California, earthquake and the hypocenter of the 1999 M 7.1
256, 1430-1431. Hector Mine, California, earthquake, J. Geophys. Res., 108,
Harris, R. A. (1998), Introduction to special section: Stress trig- 2012, doi:10.1029/2001JB000678.
gers, stress shadows, and implications for seismic hazard, J. King, G. C. P. and M. Cocco (2001), Fault interactions by elas-
Geophys. Res., 103, 24347-24358. tic stress changes: New clues from earthquake sequences, Adv.
Harris, R. A., R. W. Simpson (2002), The 1999 Mw 7.1 Hec- Geophys., 44, 1-38.
tor Mine, California, earthquake: A test of the stress shadow Marsan, D. (2003), Triggering of seismicity at short timescales
hypothesis? Bull. Seism. Soc. Am., 92, 1497-1512. following California earthquakes, J. Geophys. Res., 108, 2266,
Hauksson, E., W-C. Chi, P. and Shearer (2003), Comprehen- 10.1029/2002JB001946.
sive waveform cross-correlation of southern California seismo- Marsan, D. (2004), The role of small earthquakes in redistributing
grams: Part 1. Refined hypocenters obtained using the double- crustal elastic stress, in press in Geophys. J. Int.
difference method and tectonic implications, Eos Trans. AGU, Michael, J. and L. M. Jones (1998), Seismicity alert probabilities
84(46), Fall Meet. Suppl., Abstract S21D-0325. at Parkfield, California, revisited, Bull. Seism. Soc. Am., 88,
Helmstetter A. (2003), Is earthquake triggering driven by small 117-130.
earthquakes?, Phys. Rev. Let., 91, 058501. Mikumo, T. and T. Miyatake (1979), Earthquake sequences on
Helmstetter, A. and D. Sornette (2002), Diffusion of epicenters of a frictional fault model with non-uniform strengths and relax-
earthquake aftershock, Omori law and generalized continuous- ation times, Geophysical Journal of the Royal Astronomical
time random walk model, Phys. Rev. E., 66, 061104. Society, 59, 497-522.
X - 10HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING

Molchan, G.M. and O.E. Dmitrieva (1992), Aftershock identifi- the size of the smallest triggering event, submitted to Geophys.
cation – Methods and new approaches, Geophys. J. Int., 109, Res. Lett.
501-516. Steacy, S., D. Marsan, S. S. Nalbant and J. McCloskey
Nur A. and J.R. Booker (1972), Aftershocks caused by pore fluid (2004), Sensitivity of static stress calculations to the earth-
flow? Science, 175, 885-888. quake slip distribution, J. Geophys. Res., 109, B04303,
Ogata, Y. (1989), Statistical model for standard seismicity and doi:10.1029/2002JB002365.
detection of anomalies by residual analysis, Tectonophysics, Stein, R.S. (1999), The role of stress transfer in earthquake oc-
169, 159-174. currence, Nature, 402, 605-609.
Ogata, Y. (1992), Detection of precursory relative quiescence be- Toda, S. and R. Stein (2003), Toggling of seismicity by the
fore great earthquakes though a statistical model, J. Geophys. 1997 Kagoshima earthquake couplet: a demonstration of
Res., 97, 19845-19871. time-dependent stress transfer, J. Geophys. Res., 108, 2567,
Papazachos, B., N. Delibasis, N. Liapis, G. Moumoulidis, G. Pur- doi:10.1029/2003JB002527.
caru (1967), Aftershock sequences of some large earthquakes Utsu,T. (1969), Aftershocks and earthquake statistics (I) source
in the region of Greece, Annali di Geofisica, 20, 1-93. parameters which characterize an aftershock sequence and
Parsons, T. (2002), Global Omori law decay of triggered earth- their interrelations, J. Fac. Sci. Hokkaido Univ., Ser.VII, 3,
quakes: Large aftershocks outside the classical aftershock zone,
129-195.
J. Geophys. Res., 107, 2199.
Utsu, T., Y. Ogata and S. Matsu’ura (1995), The centenary of
Reasenberg, P. A. (1985), Second-order moment of central Cali-
the Omori formula for a decay law of aftershock activity, J.
fornia seismicity, 1969-82, J. Geophys. Res., 90, 5479-95.
Reasenberg, P. A., and Jones, L. M. (1989), Earthquake hazard Phys. Earth, 43, 1-33.
after a mainshock in California, Science, 243, 1173-1176. WGCEP (Working Group On California Earthquake Probabil-
Reasenberg, P. A., and Jones, L. M. (1994), Earthquake after- ities) (2003), Earthquake probabilities in the San Francisco
shocks: update, Science, 265, 1251-1252. Bay Region: 2002-2031, U. S. Geol. Surv. Open-File Report
Scholz, C. H. (1968), Microfractures, aftershocks, and seismicity. 03-214.
Seismological Society of America Bulletin, 58, 1117-1130. Yamanaka, Y. and K. Shimazaki (1990), Scaling relationship be-
Shaw, B.E. (1993), Generalized Omori law for aftershocks and tween the number of aftershocks and the size of the main shock,
foreshocks from a simple dynamics, Geophys. Res. Lett., 20, J. Phys. Earth, 38, 305-324.
907-910. Zeng, Y. (2001), Viscoelastic stress-triggering of the 1999 Hector
Shearer, P., E. Hauksson, G. Lin and D. Kilb (2003), Comprehen- Mine earthquake by the 1992 Landers earthquake, Geophys.
sive waveform cross-correlation of southern California seismo- Res. Let., 28, 3007-3010.
grams: Part 2. Event locations obtained using cluster analysis, Zhuang J., Y. Ogata and D. Vere-Jones (2004), Analyzing earth-
Eos Trans. AGU, 84(46), Fall Meet. Suppl., Abstract S21D- quake clustering features by using stochastic reconstruction,
0326. J. Geophys. Res., 109, B05301, doi:10.1029/2003JB002879.
Singh, S.K. and G. Suarez (1988), Regional variation in the num-
ber of aftershocks (mb ≥ 5) of large, subduction-zone earth- Agnès Helmstetter, now at Lamont-Doherty Earth Observa-
quakes (Mw ≥ 7.0), Bull. Seism. Soc. Am., 78, 230-242. tory, Columbia University (e-mail: agnes@ldeo.columbia.edu)
Solov’ev, S.L. and O.N. Solov’eva (1962), Exponential increase of
Yan Y. Kagan, Department of Earth and Space Sciences,
the total number of an earthquake’s aftershocks and the de-
University of California, Los Angeles, California. (e-mail: yka-
crease of their mean value with increasing depth, Izv. Akad.
Nauk. SSSR, seriya geofizicheskaya, 12, 1685-1694. English gan@ucla.edu)
translation 1053-1060. David D. Jackson, Department of Earth and Space Sciences,
Sornette, D. and M.J. Werner (2004), Effects of undetected seis- University of California, Los Angeles, California. (e-mail: djack-
micity on the parameters of the ETAS model: constraint on son@ucla.edu)
 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERINGX - 11

Table 1. Scaling of aftershock productivity and earthquake

frequency with magnitude. Times are in days and distances in
km. Parameters ∆m, DF , DA , NL,F , NL,A and TF are used
in the selection of mainshocks and aftershocks (see section 3.3).
Parameters tmin , tmax and λmin define the time interval used
to estimate the scaling of the number of aftershocks with the
mainshock magnitude (see section 3.4). N is the total number
of events with m ≥ md and Nms is the number of mainshocks.
K0 defined by (13) measures aftershock productivity.
# md time ∆m DF DA NL,F NL,A TF tmin tmax λmin α N Nms K0

1 2 1980-2004 1 3 2 3 2 365 0.0020 10 0.100 0.99 101402 19247 0.0092

2c 2 1980-2004 1 3 2 3 2 365 0.0020 10 0.100 1.05 101402 19247 0.0041
3c 2 1980-2004 1 3 2 3 2 36 0.0020 10 0.100 1.02 101402 33123 0.0065
4c 2 1980-2004 1 3 2 3 2 730 0.0020 10 0.100 1.06 945736 15130 0.0035
5c 2 1980-2004 0 3 2 3 2 365 0.0020 10 0.100 0.95 101402 26004 0.0167
6c 2 1980-2004 2 3 2 3 2 365 0.0020 10 0.100 1.09 101402 18784 0.0024
7c 2 1980-2004 1 3 3 3 3 365 0.0020 10 0.500 1.04 101402 19247 0.0048
8a,c 2 1980-2004 1 3 0.5 3 2 365 0.0020 10 0.005 1.16 101402 19247 0.0009
9a,c 2 1980-2004 1 3 0.5 3 2 365 0.0020 100 0.000 1.15 101402 19247 0.001
10a,c 2 1980-2004 1 3 0.5 3 2 365 0.0020 1000 0.000 1.12 101402 19247 0.0013
11c 2 1980-2004 1 1 4 1 4 365 0.0020 10 5.000 0.94 101402 37839 0.0193
12c 2 1980-2004 1 1 2 1 2 365 0.0020 10 2.000 1.01 98805 36864 0.0078
13c 2 1980-2004 1 4 1 4 1 365 0.0020 10 0.100 1.13 101402 15153 0.001
14c 2 1980-2004 1 4 4 4 4 365 0.0020 10 0.100 1.02 101402 15153 0.0064
15c 2 1980-2004 1 1 1 1 1 365 0.0020 10 0.500 1.05 101402 37839 0.003
16c 2 1980-1992.4 1 3 2 3 2 365 0.0020 10 0.100 1.00 40565 9603 0.0056
17c 2 1992.4-2004 1 3 2 3 2 365 0.0020 10 0.100 1.05 60837 9187 0.0041
18 3 1980-2004 1 3 2 3 2 365 0.0003 10 0.010 1.07 101402 19247 0.0033
19c 3 1980-2004 1 3 2 3 2 365 0.0020 10 0.010 1.10 101402 19247 0.0023
20b,c 2 1980-2004 1 3 3 4 2 36 0.0020 10 0.500 1.01 101402 29210 0.0075
21c 2 1980-2004 1 3 3 4 2 36 0.0020 10 0.500 1.02 101402 32801 0.006
X - 12HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING

3 6
10 10
3 (a) (b)
10
2 5
10 10

cumulative magnitude distribution

2
10

rate of triggered earthquakes

1 4
seismicity rate (day−1)

1 10 10
10
0 3
0 10 10
10
−1 2
−1
10 10 10

−2 −2 1
10 10 10

−3 −3 0
10 10 10

−4 −4 −1
10 10 10
−4 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7
time after mainshock (days) magnitude

Figure 1. (a) Average rate λ(mM , t) of m ≥ 3 earth-

quakes as a function of the time t after the triggering
earthquake, for different values of the magnitude mM of
the triggering earthquake increasing from 2 to 7 with a
step of 0.5 from bottom to top. The roll-over at short
times is due to the incompleteness of the catalog fol-
lowing a strong mainshock. Black lines show the fit of
λ(mM , t) by K(mM )/t0.9 in the time interval [tmin (mM ),
tmax (mM )], where tmin (mM ) = 0.0003 day (26 sec) for
mM ≤ 4.5 mainshocks. For larger mainshocks tmin (mM )
is estimated using expression (12) as the time when the
catalog is complete for m ≥ 3. The maximum time is ei-
ther 10 days or the time when the seismicity rate decays
below λmin = 0.01 day−1 . (b) Aftershock productivity
K(mM ) as a function of mM (circles) and cumulative
magnitude distribution P (m) (crosses). Solid lines are
linear regressions of K(mM ) and P (m) with exponent
respectively equal to α = 1.07 and b = 0.94.
 HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERINGX - 13

4 6
10 10
(a) (b)
5
10 3 5
10 10

cumulative magnitude distribution

4

rate of triggered earthquakes

10
2 4
seismicity rate (day−1)

3 10 10
10
2 1 3
10 10 10
1
10 0
10
2
10
0
10
−1 1
10 10
−1
10
−2 0
−2
10 10 10

−3 −3 −1
10 10 10
−4 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7
time after mainshock (days) magnitude

Figure 2. Same as Figure 1 except that we have used

md = 2 and we have corrected the seismicity rate for
missing early aftershocks using (12) (assuming GR law
with b = 1). We fit the seismicity rate in the time interval
0.002 < t < 10 days and for λ(t, mM ) > 0.1 day−1 . The
fit of K(mM ) give K0 = 0.0041 dayp−1 and α = 1.05.

5
10
4
10
3
10
seismicity rate (day )
−1

2
10
1
10
0
10
−1
10
−2
10
−3
10
−4
10
−5
10
−4 −3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10
time after mainshock (days)
Figure 3. Same as Figure 2 with different parameters
DA = 0.5 km, NL,A = 2, DF = 3 km, NL,F = 3, and
without including aftershocks in the influence zone of pre-
vious aftershocks.
X - 14HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERING

5
10

4
10

cumulative magnitude distribution

3
10

2
10

1
10

0
10

−1
10

−2
10

−3
10

2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 7

magnitude

Figure 4. Cumulative magnitude distribution of trig-

gered earthquakes, for mainshock magnitudes ranging
from 2 (below) to 7 (top) with a bin size of 0.5. The
curves have been normalized by a factor 100.5mM for clar-
ity. Dotted lines represent a GR law with b = 1 for
reference. The number of events decreases with mM (be-
tween N = 11502 for 7 ≤ mM < 7.5 down to N = 1328
for 7 ≤ mM < 2.5) due to the smaller fraction of small
earthquakes considered as mainshocks, and due to the
shorter time window used for aftershock selection of small
mainshocks.

−1 (a) D=1 −1 (b) D=1

10 10

−2 −2
10 10
pdf

pdf

−3 −3
10 D=2 10 D=2

−4 −4
10 10

−5 −5
10 −2 −1 0 1 2 3
10 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
interevent distance (km) interevent distance (km)

Figure 5. Distribution of distances between hypocenters

pr (r, t) for SHLK (a) and HCS (b) catalogs, using only
earthquake pairs with inter-event times in the range [t,
1.25t], where t increases between 1.4 minutes (blue curve)
to 2500 days (red curve). The fractal dimension of pr (r),
measured for 0.1 ≤ r ≤ 5 km for all earthquakes (black
lines), is D = 1.54 for SHLK catalog and D = 1.73 for
HCS catalog.
HELMSTETTER, KAGAN, AND JACKSON: STRESS TRANSFERS AND EARTHQUAKE TRIGGERINGX - 15

2
hypocenters correlation dimension

1.5

0.5

0
−3 −2 −1 0 1 2 3
10 10 10 10 10 10 10
interevent time (days)

Figure 6. Fractal dimension of pr (r, t) as a function of

inter-event time t, using SHLK catalog (diamonds) and
HCS catalog (circles).