Delayed and Sustained Remote Triggering of Small Earthquakes in the San Jacinto Fault Region by the 2014 Mw 7.2 Papanoa, Mexico Earthquake

We apply a move max matched‐filter method to detect heightened seismicity triggered in the San Jacinto fault (SJF) zone, by the 2014 Mw 7.2 Papanoa, Mexico earthquake. The move max matched filter detects 5.4 and 1.7 times the number of earthquakes in the Southern California Seismic Network catalog and those detected by the matched‐filter method, respectively. The seismicity rate increases significantly ~3.5 hr after the passage of the teleseismic waves and persists above the background level for about 1 week. This observation of delayed triggering may imply that dynamic stresses had initiated a time‐dependent advance to failure or a secondary process. A highly active triggered patch is located ~10 km west of the SJF near Anza, on a previously unmapped fault. Focal mechanisms and a best fit plane suggest a normal fault perpendicular to the SJF. The unmapped fault may indicate higher seismic hazard in the surrounding areas if a large earthquake nucleated around the Anza seismic gap.


Introduction
With the implementation of globally distributed seismic networks and improved earthquake detection/location methods, large earthquakes have been widely observed to change the seismicity rate both dynamically (instantaneously) and in a delayed manner, within regions hundreds to thousands of kilometers away (Ghosh et al., 2009;Hill et al., 1993;Mendoza et al., 2016;Meng & Peng, 2014;Shelly et al., 2011). The dynamic stress changes due to large earthquakes at teleseismic distances have been estimated to be between 0.1 and 1 MPa (Prejean & Hill, 2009); thus, the occurrence of triggered earthquakes indicates that faults in that region are sensitive to small stress perturbations. For this reason, studies of triggered earthquakes provide clues to understanding the state of stress of critically stressed and potentially seismically hazardous faults (Johnson et al., 2015).
The San Jacinto Fault Zone (SJFZ) is currently the most seismically active fault segment in southern California (Hauksson et al., 2012;Kagan et al., 2006). It consists of several right lateral strike slip faults that -have produced 11 earthquakes with magnitudes larger than 6 in the last 120 years (Bailey et al., 2010;Kagan et al., 2006). The central region of the SJFZ, a segment known as the Anza seismic gap, has not spawned a major earthquake for over 200 years (Rockwell et al., 1990). Previous studies show that there is a potential for an earthquake with magnitude 6.5 or larger to occur in this seismically quiet segment (Sanders & Kanamori, 1984;Thatcher et al., 1975). Therefore, understanding the stress state and the conditions required for rupture nucleation in the Anza gap or on neighboring faults is important for estimating the earthquake hazard in this area.
At 14:27:24 (UTC) on 18 April 2014, the Mw 7.2 Papanoa, Mexico earthquake occurred within the Guerrero seismic gap, as a result of thrust motion at shallow depth along the Cocos North America plate boundary (Mendoza & Martínez López, 2017). The event resulted in unprecedented damage near the epicenter and minor damage in Mexico City. In this study, we build upon the commonly used template matched lter method and apply the move max matched lter (MMMF) method to detect seismic events -fi -fi in this region, 1 month before and after the Papanoa earthquake. A similar method has been applied previously by Shelly et al. (2007) to detect weak low frequency earthquakes. We then use hypoinverse " " -(Klein, 2002) and hypoDD (Waldhauser, 2001) to locate and relocate all detected events with high precision. Finally, we investigate the spatiotemporal distribution of relocated earthquakes, compare them to other earthquake catalogs, and discuss the possibility that these events were triggered and, if so, investigate the triggering mechanisms.
©2019. The Authors. This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided the original work is properly cited.

10.1029/2019GL084604
Key Points: • -The move max matched filter method shows higher capability than traditional matched lter -fi method in detecting small earthquakes • More than one physical mechanism can explain the delayed and sustained triggering of small earthquakes in the San Jacinto Fault region • The unmapped fault distinguis hed in this study may indicate more serious seismic hazards in the Anza region Supporting Information: • Supporting Information S1 • Figure S1 • Figure S2 • Figure S3 • Table S1 • Table S2 • Table S3 • Table S4 • Data Set S1

Data and Methods
In this study, we use continuous seismic data from 17 stations surrounding the San Jacinto fault (Figure 1) to detect and locate local earthquakes in the study region that were triggered by the 2014 Mw 7.2 Papanoa earthquake. The data spans 1 month before and 1 month after the Papanoa event. During this period, there are 880 local events recorded by the Southern California Seismic Network (SCSN) catalog ( Figure 1). All events in this catalog are used as templates to perform the matched lter and MMMF methods in detecting -fi any local event missed (or undetected) by the network. For each catalog (template) event, we cut a 20 s time window (resampled at 20 samples per second) after the event origin time. Afterward, the template is used to perform a sliding time window search throughout the 2 to 7 Hz band passed data set, (with 0.05 s time ----steps) to detect other similar or correlated waveforms.
We rst use the traditional template matched fi -filter method (Gibbons & Ringdal, 2006). This method is useful in developing larger catalogs, since it works on a well tested assumption: If the template and matched wave-form are highly correlated, the matched waveform represents an event that shares a similar location, station move-out, and focal mechanism as the template event. The method calculates the cross-correlation coefficients between the template event and the continuous waveforms in each of the three station channel components, sums the coef cients across all stations, and retains newly detected events that met or exceed some predeterfi mined, con dence threshold value. After experimenting with several thresholds by visually checking the seisfi mograms of randomly selected events detected by the traditional template matched-filter method, we chose a uniform threshold value of 6 times the root mean square to detect new events in the 2 month data period. - One signi cant limitation of the traditional matched lter method, though, is that any given template does fi -fi not usually produce the maximum cross correlation coef cient at the exact same move out time in all sta--fi tions for the missing events. Furthermore, for the same time step, some stations show peak coef cients, fi while others show much smaller or even negative coef cients, with peak coef cients a few time steps fi fi (0.05 1 s in our case) away, which lowers the summed coef cients under the threshold (supporting informa--fi tion Figure S1a). This is because even though some events are located close to the catalog events, their spatial offset is signi cant enough to result in different time lags across each station, relative to the template event.
fi To reconcile this, Zhang and Wen (2015) developed the Match and Locate method to improve their detections by adding potential hypocentral locations on grid nodes around the template location. They then calculate the theoretical move out at each potential location and time shift the seismic waveform accordingly before stacking and calculating the cross correlation coef cients. The Match and Locate method can detect fi small events near the template location but may still miss events that lie in areas between the template location and the grid nodes. In addition, the velocity model used to compute the traveltime may be crude and result in an inaccurate move out calculation. Furthermore, it is computationally inef fi cient to calculate the move out for all the grid nodes to all the stations. The MMMF method implemented with the --SEC C code (Senobari et al., 2018), we use in this study builds upon these traditional methods, yet overcomes their limitations. The rst step of the MMMF method is similar to the traditional method by calculating the fi cross coef cient for each station channel component; however, instead of summing the cross coef cients fi --fi directly with constant move out for each template, the MMMF method dynamically replaces the coef cient fi with the maximum coef cient 1 s before and after the current time step (Figures S1b and S1c). In other fi words, we allow discrepancies in the locations within a 1 s traveltime distance away from the template event. Our method then sums the move max coef cients and removes the mean values to make it comparfi able with the matched lter results ( Figure S1d). This method applies the same threshold used in the -fi matched lter method to obtain earthquake detections. -fi After obtaining a detection catalog, we use the cross correlation between the template and the matched -event for each station to calculate and wave arrival times on the vertical and horizontal components, P S respectively. Lastly, we use a 1 D velocity model from Scott et al. (1994) to locate all detected and phases -P S using hypoinverse (Klein, 2002) and then relatively relocate them with higher precision using hypoDD (Waldhauser, 2001).

Results
Local earthquakes in the San Jacinto fault region triggered by the Mw 7.2 Papanoa earthquake are shown in Figures 2b and 2c. These gures show a remarkable increase in microseismic activity after the Papanoa fi earthquake, particularly, a few hours after the passage of the teleseismic wave train. This heightened seismic activity is also shown in a spectrogram and histogram (Figures 2c and 3a). Using the 880 catalog events in the 2 month data period as templates, the matched lter and MMMF detect 2,835 and 4,765 local earthquakes --fi  Table S1. The red star in the inset map shows the location of the Papanoa earthquake, while the red square marks the study area presented here. Three beach balls determined in this study show normal fault focal mechanisms in the off fault earthquake cluster discussed in the main text. The black dashed arrow indicates the great circle travel path of the teleseismic (Papanoa) event to the study region. (b). Distribution of detected earthquakes. The blue dots represent earthquakes detected by the move max matched lter method and located using hypoDD, while the -fi red dots represent earthquakes from the Quake Template Matching catalog (Ross et al., 2019). The magenta and yellow polygons encompass the relocated earthquakes used to show the cross section along Lines A B in Figure S3 and C D in Figure 4b. (c). Seismicity rate ratio plot (after and before the Mw 7.2 Papanoa --earthquake). Since the cutoff here is selected to be 1, the colored region represents the area of seismicity rate increase. The red ellipse shows the most seismically active region that we interpreted to be triggered by the Papanoa earthquake. The red square marks the town of Anza, California.
about 3.2 and 5.4 times the number of catalogued events, respectively (Figure 3a). These results show an elevated level of seismicity that lasts up to 7 days before returning to the background level. To better assess how the newly detected events compare with the original catalog events, we normalize the cumulative  detections ( Figure 3a). All three lines in Figure 3a show similar cumulative patterns, yet differ in detail. In the SCSN catalog, there is an~50% increase in the number of events occurring 1 month after the Papanoa earthquake, while the matched lter detection results show an~87% increase in detected events occurring -fi after the Papanoa earthquake during the same period. More interestingly, our MMMF method detects the most events and also shows a~50% increase in local seismicity as seen in the catalog events, amplifying the detection nearly the same level both before and after the Papanoa earthquake but additionally showing many more detected events in the rst 3 days following the mainshock (Figure 3a). fi Although the relocated 3,034 local events show an overall similar distribution to that of the catalog events, we note a few signi cant differences found in detail. The relocated MMMF catalog reveals some linear feafi tures along the fault traces (Figure 1b), as well as a sparse area of seismicity between the northwest and southeast earthquake clusters. This less active seismic region is de ned as the Anza gap (Rockwell et al., fi 1990;Sanders & Kanamori, 1984;Thatcher et al., 1975). This gap lies in close proximity to the seismically active triple junction of the Coyote Creek, Clark, and Buck Ridge strands of the SJFZ. A cross section along the SJF trace shows that the earthquake depth distribution is shallow when approaching the Anza gap from the northwest, but then increases in depth when moving away from it towards the southeast ( Figure S3).
To investigate the seismicity rate changes in detail, we divide the study region into 0.025°by 0.025°grids and compare their seismic activity 1 month before and after the Mw 7.2 Papanoa earthquake. Figure 1c shows a general seismicity rate increase after the Papanoa earthquake in the study region. Interestingly, new patches of seismicity are observed northwest and southeast of the Anza gap, but no heightened seismicity is observed within the Anza gap itself. In fact, the patch with the highest seismicity rate increase is located~10 km west of the town of Anza and off the SJF where there are no mapped faults. Since the earthquake cluster shows a planar distribution in 3 D, we determine a best t plane that has a strike of 215°from -fi north nearly perpendicular to the strike of the SJF, which is about 306°and is steeply dipping 83°to -the northwest (Figure 4).

Discussion
Using the 880 catalog events as templates, the traditional matched lter method detects 2,835 events, -fi whereas the MMMF detects 4,765 events approximately 1.7 times that of the matched lter detections.
--fi This result indicates that the MMMF method has a much higher detection capability compared to traditional template matching. We note that the Quake Template Matching (QTM) catalog (Ross et al., 2019), which is created by the template matching method, shows higher earthquake detections (5,934) for the same period. This is because a shorter template window ( 2.5 s for wave templates and 4 s for wave templates) is used, ≤ P S while here, we use a 20 s time window for both and wave templates. This could result in some smaller -P S magnitude events being left out of our catalog since they fell below out pre set detection thresholds, but for each template event, the MMMF method can detect events within a certain distance range away from the template event, which depends on the allowed time shift (1 s in our case), the velocity model, and station event distances. Thus, in the QTM catalog, all the earthquakes are in close proximity to the template events (Figure 1b), but our MMMF catalog shows earthquakes distributed over a broader area relative to the template location. Despite differences in the methodology used to derive the QTM and MMMF catalogs, the fault traces delineated are fairly similar. As shown in Figure 1b, both catalogs show seismicity that is spatially distributed around the mapped faults that trend NW SE. Both catalogs also show that seismicity extends laterally a few kilometers on both sides of the fault traces (Figure 1), down to a depth of 20 km (Figures 4 and S3), with some cross cutting structures in between. However, at ner scales, it becomes fi increasingly dif cult to interpret new faults traces, especially for small structures where there are limited fi earthquakes. In the MMMF catalog, we can better capture the unmapped fault by interpreting a broader plane (Figure 4). The MMMF method is also more computationally ef cient than the traditional template fi matching method, especially for large catalogs. It does not require all events recorded in the catalog to be used as templates, as opposed to traditional template matching methods. The study region can be divided into many grid areas based on earthquake distribution for further analyses. When allowing the crosscorrelation coef cients to move up to certain time steps, only a few template events depending on the comfi plexity of local structures and similarity of the waveforms are needed to capture nearly any missing event in a given grid area.
The seismicity rate changes nonuniformly in the study region, with the most signi cant patch of seismicity fi about 10 km west of the SJF and near the town of Anza. Figure 4 shows the 3 D distribution of the relocated earthquakes in this small region, de ning a planar structure between depths of 3.5 to 11.5 km that is perpenfi dicular to the SJF. A best t plane for these locations shows a strike of 215°that is nearly perpendicular to the fi SJF (306°) and dips steeply at 83°to the northwest (Figure 4). The focal mechanisms of the three largest magnitude (SCSN catalog) earthquakes (Mw 2.53, 1.73, and 2.13) within this patch indicate a normal fault with a strike also perpendicular to the SJF and is steeply dipping (Table S3). This unmapped fault could be the result of the extensional stress from the right lateral San Jacinto and Elsinore fault. A fault of this length (~2.5 km) and width (~8 km) could produce an earthquake up to magnitude 5.3 (Wells & Coppersmith, 1994).
We next investigate the events on this blind fault area recorded in the SCSN catalog for the past 17 years (1 January 2001 to 30 December 2017). We nd very little and sparsely distributed seismicity recorded in the fi catalog before the Papanoa earthquake, implying that this fault has been in a state of seismic quiescence. This quiescence may be due to low tectonic shear stress loading it, since it is nearly perpendicular to the overall kinematic motion of the plate boundary in this region. In addition, the magnitude evolution plot (Figure 3b) using data from the SCSN and QTM catalog (Ross et al., 2019) shows the largest magnitude event in this small region to be a Mw 2.53 that occurred~5.86 hr after the teleseismic event. In other words, it occurred after the initiation of heightened seismicity with smaller magnitudes and demonstrates that this delayed triggering instance is not a classic example of a mainshock aftershock sequence. Because this fault did not produce a similar burst of seismicity in the last 17 years, we infer that the burst of seismicity observed in this region is likely delayed triggered caused by the Mw 7.2 Papanoa earthquake and unlikely to be a random (background) occurrence.  Figure 1c. (b) Cross section along the Line C D, which is perpendicular to the strike of the best t plane fi (Table S3). Earthquakes in the yellow polygon (Figure 1b) are plotted. The red dashed line marks the dip of the best t fi plane. The cross section is true scale, and plotted below mean sea level. This unmapped fault does not seem to extend long enough in its strike direction and is~10 and~35 km away from the SJFZ and the Elsinore Fault Zone, respectively. Thus, it limits the likelihood of a through going rupture from one fault zone to the other (Harris & Day, 1993;Oglesby, 2005), which could increase the potential for a large earthquake in this region. However, it is also possible that the fault could be larger, but only a section of it is active during the time period of this study. In addition, this newly observed normal fault is in close proximity to the Anza gap, an area that has the potential to produce a Mw 6.5 or larger earthquake in the near future (Sanders & Kanamori, 1984;Thatcher et al., 1975). If a large enough earthquake were to occur in the Anza gap, it could trigger earthquakes on this unmapped fault and increase the seismic hazard in the region. The July 4th, 2019 Mw 6.4 Searles Valley and the July 6th, 2019 Mw 7.1 Ridgecrest earthquakes are a sobering example of the hazards posed by unmapped faults in California (Rollins et al., 2019).
In this study, we do not observe dynamically (instantaneously) triggered earthquakes during the passage of the seismic wave train of the 2014 Papanoa earthquake. Still,~3.5 hr after the passage of the mainshock coda waves, we observe a burst of small local earthquakes. Prior to the arrival of the teleseismic event, the background seismicity rate in the SCSN, QTM, and MMMF catalog is about 9, 75, and 66.4 events per day, respectively. The seismicity rate increased signi cantly to 79, 583, and 408 events per day in the fi following 24 hr after the passage of the teleseismic event (Table S2), with heightened seismicity lasting for about 1 week before returning to ambient seismicity levels. This observation may suggest that dynamic stresses imparted by the passage of teleseismic waves may have invoked a time dependent advance to fail-ure process or other secondary mechanisms that resulted in delayed triggering (Freed, 2005). In the former, the dynamic stresses can in uence a fault's earthquake cycle by time advancing it to failure. In other words, fl faults closer to failure will advance further in time to failure, than faults further away in time from failure (Gomberg et al., 2001;Mendoza et al., 2016). In a population of faults, there will usually be a small fraction of them that are exceptionally close to failure, whereupon the addition of any small stress perturbation can lead to the occurrence of brittle failure (Brodsky & van der Elst, 2014;Gomberg et al., 1998). One proposed mechanism that supports this, is the subcritical crack growth model, which suggests that the intensity of stress at the crack tip (proportional to crack size)controls the rate of crack growth i.e. earthquake nuclea-tion (Atkinson, 1984). The crack extends very slowly at the start, but when there is a sudden increase in stress at the crack tip, the crack quickly grows, thus leading towhat is an apparent delayed fault rupture (Atkinson, 1984;Rinne, 2008). For the seismicity in the unmapped fault region, we observe that the time interval between two consecutive events decreases until there is a burst of seismicity ( Figure S4a). In addition, the area of seismicity on this blind fault extends from southwest to northeast and may suggest growth of this fault (Figures 4b and 4c).
Another possible mechanism for delayed triggering in this region could be the enhancement of permeability and subsequent increase in pore pressure due to the oscillatory stresses initiated by teleseismic waves. When the dilatation component of compressional waves interacts with uids in the crust, it can rupture uid seals fl fl within the fault, unclog fractures, remove temporary barriers, and produce subsequent uid ow leading to fl fl an increase in pore pressure (Brodsky & van der Elst, 2014;Hill et al., 1993). Previous studies show that permeability enhancement increases with amplitude of the seismic waves in Pinon Flat, California, and sustained water level changes at Long Valley Caldera, California, are associated with remote earthquakes up to several thousand miles away (Brodsky et al., 2003;Elkhoury et al., 2006;Roeloffs et al., 2003).
An alternative class of mechanisms to explain the delayed triggering phenomenon in the San Jacinto fault region is that the Papanoa earthquake may have rst triggered a local creep or slow slip event (SSE) that then fi triggered nearby local earthquakes. Using seismic and geodetic data, Wdowinski (2009) proposed that deep creep causes excess seismicity along the San Jacinto fault. Shelly et al. (2011) show that fault creep and triggered tremor initiated by remote earthquakes can last for several days after the passage of teleseismic waves, resulting in a prolonged increase of local seismicity. Other studies have shown that SSEs can trigger both small and large earthquakes and sustain this process as the SSE evolves in space and time (Delahaye et al., 2009;Gomberg et al., 2012;Kato et al., 2012;Peng & Gomberg, 2010;Vidale et al., 2011). Radiguet et al. (2016 infer that an SSE in Guerrero, Mexico, triggered the 2014 Mw 7.2 Papanoa earthquake. Unfortunately, nearby strainmeter stations are unable to identify evidence of fault creep or SSE in the region during this period of elevated seismicity. This may be because stations are not close enough especially to the most activated unmapped fault area or perhaps slip is too small to be detectable. Still, recent studies have identi ed tectonic tremors in the study region (Hutchison & Ghosh, 2017;Wang et al., 2013). Even fi though there are observations that SSEs and tremors occur independently (Delahaye et al., 2009;Hutchison & Ghosh, 2016;Li & Ghosh, 2017), generally, they are associated with each other and are known as episodic tremor and slip (Ghosh et al., 2012(Ghosh et al., , 2015Ito et al., 2007). Tremors and slow slip are most likely the result of the same physical processes along the fault plane (e.g., Ghosh et al., 2012). Thus, a triggered SSE or fault creep in the study region may be responsible for the delayed and sustained increase of seismicity around the SJF. Nonetheless, it is dif cult to determine which mechanisms, and to what extent, are driving fi delayed triggering in the study region.

Conclusions
In this study, we use the MMMF method to capture delayed triggering of small local earthquakes in the SJFZ near the Anza Gap in California. The MMMF method detects about 5.4 times the amount of local seismicity found in the SCSN catalogs and 1.7 times the amount of seismicity identi ed by the traditional matched fi filter method, thus greatly improving the completeness of the current earthquake catalog. Although there is no observation of dynamically triggered events during the passage of the teleseismic surface waves, we nd fi a signi cant increase in the seismicity rate that begins a few hours after the coda and persists for about fi 1 week, before decaying back down to the background seismicity level. More than one triggering mechanism could be responsible for these observations. They include, but are not limited to: (1) dynamic stresses initiating a time dependent, advance to brittle failure process, (2) a transient increase in pore pressure, or (3) sec-ondary mechanisms such as fault creep or a SSE. Relocations of seismicity detected by the MMMF method show several linear clusters of seismicity occurring along the SJF system and reveal depth variation, where seismicity is shallow northwest of the Anza gap, and then become increasingly deep as the locations migrate to the southeast. Lastly, a highly active cluster of triggered seismicity implies the existence of an unmapped blind fault that is trending nearly perpendicular (and in close proximity) to the SJF. This blind fault is located ~10 km west of the Anza gap and the town of Anza. If coseismically triggered, it may contribute to the high seismic hazard already posed by the southern San Andreas fault where a large earthquake is overdue. This work is funded by the NSF Award EAR 1358686 and 1620655. The author thanks the European Union's Horizon 2020 Programme under the ChEESE Project (https://cheese coe.eu/), grant agreement no. 823844, for partially funding this work. The continuous seismic data used in this study is provided by the Incorporated Research Institutions. The authors would like to thank members of the Ghosh Earthquake Seismology Lab, David D. Oglesby and Gareth Funning for their insightful suggesti ons. We also want to extend our thanks to Debi Kilb, an anonymous reviewer, and the editor Gavin Hayes for their constructive comments that helped improve the content of this manuscript.