Brittle-failure earthquakes in the lower crust, where high pressures and temperatures would typically promote ductile deformation, are relatively rare but occasionally observed beneath active volcanic centers. Where they occur, these earthquakes provide a rare opportunity to observe volcanic processes in the lower crust, such as fluid injection and migration, which may induce brittle faulting under these conditions. Here, we examine recent short-duration earthquake swarms deep beneath the southwestern margin of Long Valley Caldera, near Mammoth Mountain. We focus in particular on a swarm that occurred September 29–30, 2009. To maximally illuminate the spatial-temporal progression, we supplement catalog events by detecting additional small events with similar waveforms in the continuous data, achieving up to a 10-fold increase in the number of locatable events. We then relocate all events, using cross-correlation and a double-difference algorithm. We find that the 2009 swarm exhibits systematically decelerating upward migration, with hypocenters shallowing from 21 to 19 km depth over approximately 12 hours. This relatively high migration rate, combined with a modest maximum magnitude of 1.4 in this swarm, suggests the trigger might be ascending CO2 released from underlying magma.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 Our objective in this study is to develop high-resolution images of three swarms of brittle-failure earthquakes that occurred in the lower crust beneath Mammoth Mountain in 2006, 2008, and 2009, as a step toward illuminating ongoing processes in the roots of the Mammoth Mountain mafic magmatic field. Our focus in this paper is primarily on the 2009 swarm as it is the best recorded of the three.
1.1. Tectonic Setting
 Mammoth Mountain is a cluster of dacitic domes erupted between 100 to 50 ka standing on the southwest topographic rim of Long Valley caldera in eastern California. Long Valley caldera was formed 760,000 years ago with the massive eruption of 600 km3 of rhyolitic magma in the form of the Bishop tuff [Hildreth, 2004]. The Long Valley volcanic field (LVVF), which includes Mammoth Mountain and the Inyo-Mono volcanic chain extending 45 km north of Mammoth Mountain, has been the dominant source of eastern California volcanism over the past 2 Ma. Focusing of volcanism in the LVVF is likely related to 1) its location at the west end of the Mina deflection, a zone of sinistral strike-slip faults forming a right-stepping (releasing) offset between the north end of the dextral Eastern California Shear Zone with the dextral faults of the Walker Lane [Hill, 2006], and 2) its position above the westward-propagating delamination of the dense lithosphere beneath the Sierra Nevada, exposing the base of the crust to hot asthenosphere [Frassetto et al., 2011]. Mammoth Mountain itself is surrounded by a field of monogenetic mafic volcanic vents that erupted basaltic magma between 200 and 8 ka. Red Cones, located 3 km SSW of Mammoth Mountain, is the youngest of these vents. Long Valley caldera experienced a two-decade period of volcanic unrest between 1980 and 2000, which included recurring, energetic earthquake swarms (the largest included four M 6 earthquakes near the caldera in May 1980) and cumulative uplift of the resurgent dome in the center of the caldera by nearly 80 cm [Hill, 2006]. Mammoth Mountain participated in this unrest with an 11-month-long earthquake swarm in 1989–90 that was followed by the onset of diffuse magmatic CO2 degassing and mid-crustal long-period (LP) volcanic earthquakes, which continue to this time [Hill and Prejean, 2005].
1.2. Earthquakes Beneath Mammoth Mountain
Figure 1 illustrates the zones and modes of seismicity beneath Mammoth Mountain. Shallow, brittle-failure seismicity in the vicinity of Mammoth Mountain is active and mostly limited to the upper 8 km [Hill, 1992]. Small, shallow earthquakes often occur in swarms, sometimes as rapid sequences of overlapping events, termed spasmodic bursts. In the mid-crust, 10–18 km depth, brittle-failure events are mostly absent, replaced by long-period (LP) events that probably reflect fluid movement in the ductile regime [Pitt et al., 2002; Hill and Prejean, 2005].
 Below the LP events, high frequency brittle failure events resume, mostly occurring in short duration swarms. Lower crustal swarms occurred in June 2006 and September 2009, with a smaller swarm in January 2008. All swarms exhibited relatively abrupt onset and termination, with the vast majority of events occurring within less than 24 hours. Epicenters were concentrated 2–4 km NW of Red Cones, 3–6 km SW of the Mammoth Mountain summit (Figure 1). The 2006 swarm (53 events in the NCSN catalog, maximum magnitude 2.3) was the deepest, at 26–31 km, approaching the Moho depth of ∼35 km [Frassetto et al., 2011]. The 2008 (8 catalog events, M ≤ 1.6) and 2009 (55 catalog events, M ≤ 1.4) swarms were shallower, mostly between 19 and 22 km depth.
1.3. Other Observations of Deep Brittle-Failure Earthquakes
 The Lake Tahoe swarm occurred ∼200 km NW of Long Valley and lasted about 6 months, beginning in August 2003 [Smith et al., 2004; von Seggern et al., 2008]. During the first few weeks of the swarm, hypocenters migrated upwards from 33 to 29 km depth, at an average rate of 2.4 mm/s (∼200 m/day), producing earthquakes up to M 2.2. Continuous GPS data revealed that the swarm was accompanied by 8 mm of uplift and 6 mm horizontal displacement at one nearby site.
 Activity at the northern volcanic rift zone of Iceland also includes migrating swarms of brittle failure events below the typical brittle-ductile transition. Using a dense temporary seismic deployment, White et al.  observed brittle failure earthquakes at depths of 13.5–17.5 km. Short duration bursts lasting on the order of hours propagated both downward and upward along the inferred dike surface, at velocities of ∼4 cm/s. In addition to migration of short-duration bursts, a slower, longer-term upward migration of a few km in a few weeks was also observed, similar to the Tahoe swarm.
2. Data and Methods
 Our approach to data analysis for the Mammoth Mountain swarms involves two steps: 1) utilizing the waveforms of earthquakes included in the Northern California Seismic Network (NCSN) catalog to identify similar small earthquakes not included in the catalog and 2) precisely relocating both cataloged and newly-detected events in a consistent framework. These objectives are accomplished in tandem, with detection and relocation relying on cross-correlation measurements between similar events at the same station. Precise correlation-derived differential times are finally input into a double-difference relocation algorithm (hypoDD) [Waldhauser and Ellsworth, 2000].
 NCSN catalog events and their hypocenters are based on routinely processed data from seismic stations comprising the Long Valley Caldera seismic network, a dense node within the NCSN. Most Long Valley stations have short-period (1 sec) single-component vertical seismometers, although as of 2006 there were three 2-component (vertical plus one horizontal) and four 3-component stations within 20 km of Mammoth Mountain. An additional 3-component broadband station (MDPB, located near Devil's Postpile ∼4 km west of Mammoth) was operational by the time of the 2009 swarm.
 In this study, we use NCSN catalog events as “waveform templates.” Using these template events enables us to identify low signal-to-noise events with similar waveform timing and shape in the continuous seismic data [e.g., Shelly et al., 2007]. For each cataloged earthquake, we construct separate P and S-wave templates for each station with a P arrival time in the NCSN catalog phase data. Each template is 2.5 seconds in length, beginning 0.5 seconds prior to the P or S arrival time, and waveform data is filtered 2–10 Hz. For stations with no S arrival time we estimate the arrival time based on the catalog origin time, the P-wave arrival time, and an assumed P to S velocity ratio (vp/vs) of 1.75. While most stations are vertical component only, we use all available components, forming a P and S template for each component. While in theory stations above deep sources will have very little S-wave energy on the vertical component, in practice verticals commonly record a clear S-wave, probably as a result of a near-surface S to P conversion. Multi-component stations in the area support this relationship. The use of converted waves should not bias our location results, since we only use relative (not absolute) arrival times (see below).
 After constructing the waveform templates, we then scan them through the continuous seismic data at each corresponding station and channel, measuring the correlation coefficient at every lag time, incrementing by 0.005 seconds. This interval is smaller than necessary for event detection, but it is useful for later event location. Events need not be identical to be detected – we simply search for events that correlate at levels above the background. Because of waveform polarity variations probably resulting from diversity among earthquake focal mechanisms, we work with the absolute value of the correlation coefficient, rather than the coefficient itself. Thus, strongly negative correlations are considered equally with strongly positive correlations. By doing this, we increase susceptibility to cycle slips (finding a maximum correlation one waveform cycle, or in this case, a half cycle, from the true alignment); however, we gain the potential ability to obtain precise relative locations between events with different focal mechanisms.
 Initial event detection is determined by the sum of absolute correlation coefficients (P and S) at all stations at each lag time. For each day of absolute correlation coefficient sums xi, we use a detection threshold of threshi = 8 * MAD(xi) + median(xi), where MAD is the median absolute deviation. Those events that exceed this threshold are selected for further examination. For these events, we then attempt to measure the differential times at each seismic channel between the template event and events detected in the continuous data (often including other cataloged events). We allow a maximum time differential of 0.15 s for P and 0.25 s for S, relative to the original template, which allows for event separation on the order of a few km, and we require that correlations exceed 7 * MAD(xj) + median(xj), where xj is the daily vector of absolute correlation coefficients for each data channel. On average, this scheme produces a correlation threshold near 0.6; however, we prefer this method to a fixed correlation threshold, since some data channels have much larger background fluctuations in the correlations than others. These differential times are then directly input into a double-difference location algorithm (hypoDD) [Waldhauser and Ellsworth, 2000], weighted by the squares of their correlation coefficients. We also incorporate differential times based on the catalog P and (where available) S arrival times, constraining the relative locations of events that are not well tied with cross-correlation differential times. An important strategy in the inversion is that we initially down-weight the correlation-derived differential times to establish a framework of locations based on the catalog differential times. Subsequently, we reverse this weighting scheme, refining locations in the last several iterations based on the correlation differential times. Finally, we keep those events that retain at least 4 P-wave and 4 S-wave cross-correlation differential times through the inversion (i.e., they are not discarded as outliers).
Figure 1 shows seismicity beneath Mammoth Mountain, including the deep 2006, 2008, and 2009 swarms. Figure 2 shows the space-time progression of the swarms. Compared with the other sequences, the 2006 swarm is both the deepest and the most diffuse, although many of the events are members of a cluster near 28.5 km depth. The 2006 swarm also has a lower degree of waveform similarity between events, perhaps in part because of its larger spatial extent. This results in more modest gains from the correlation-based detection and location scheme than for the 2009 swarm. For 2006, we obtain 140 locatable events starting from 53 events in the NCSN catalog. The smaller 2008 swarm occurs shallower than the 2006 swarm, and seems to show upward migration from 22–21 km. For this swarm, we obtain 21 locatable events starting from 8 catalog events.
 Except for a few deep outliers, the 2009 swarm seems to begin where the 2008 swarm left off – it starts at ∼21 km depth and propagates upward to ∼19 km depth (Figures 2b and 3). Because of the high degree of waveform similarity in the 2009 swarm, the correlation-based detection/location method is highly successful, resulting in 602 locatable events (starting from 55 cataloged events). Together, these events show systematically decelerating upward migration. While propagation of the early part of the swarm could be approximated by a diffusional model (where distance is proportional to the square root of time), upward migration plateaus at ∼19.5 km depth.
3.2. Waveform Character
Figure 3 shows the waveforms for the cataloged events from all three swarms as recorded at the Red Cones station (MRD), located almost directly above the seismicity. Gradually decreasing differential S and P arrival (S-P) times confirm the systematic upward migration during the 2008 and 2009 swarms. The 2006 events have longer S-P times, consistent with their deeper source locations. At MRD, the S waves appear much weaker in the 2006 swarm than those in the 2008 and 2009 swarms, despite having similar epicenters, but this is not true on all stations (Figure S3 of the auxiliary material). This could imply either high attenuation directly above the 2006 swarm (on the path to MRD) or alternatively a different distribution of focal mechanisms of the earthquakes. A few of the 2006 events seem to have an extended, long-period coda.
 The 2009 swarm began with an earthquake that was relatively enriched in LP energy compared to the majority of 2008 and 2009 events, and at least 2 events with similarly enhanced LP energy occurred later in this sequence (Figure 3). These characteristics suggest that fluid movement may have accompanied brittle faulting for these events. In particular, the LP-like character of the first event of the 2009 swarm suggests that swarm initiation might have corresponded with the abrupt release of high-pressure fluid.
3.3. P-Wave Polarities
 We examined P-wave first motions in an attempt to constrain the focal mechanisms of the earthquakes. A strongly one-sided station distribution prevents obtaining well-constrained mechanisms. However, diversity in mechanisms is clear, as earthquakes with nearly the same location exhibit very different first motion patterns, sometimes appearing completely reversed from an event just a short time later (Figure S2 of the auxiliary material). Figure 3 (inset) highlights the variability in P-wave first motions at station MRD.
 The mechanism or mechanisms by which brittle-failure lower-crustal earthquakes occur is a matter of debate, since the lower crust is thought to typically be too warm to host earthquakes, instead accommodating deformation through ductile flow [Sibson, 1982]. Most proposed mechanisms implicate fluids in some manner, either as promoting failure by reduction of normal stress on pre-existing fractures [Reyners et al., 2007] or by inducing high strain rates in the surrounding rock [e.g., Soosalu et al., 2010]. Possible fluids range from magma to CO2 to water, all of which may be present in a volcanic system. Earthquakes may reflect either hydrofracturing (if the fluid pressure exceeds the least principle stress) or slip on pre-existing fractures triggered by the reduction in effective normal stress resulting from an increase in fluid pressure. A migrating earthquake sequence could be triggered by fluid diffusion extending the full distance of the earthquakes, or by a migrating fluid pressure pulse in an already pressurized system.
 Differences in migration rates among swarms may relate to the properties of the triggering fluid, especially its viscosity [Rubin, 1995; Waite and Smith, 2002]. Compared with the 2009 Mammoth swarm, the 2003 Tahoe swarm [Smith et al., 2004] was deeper, longer lasting, and more slowly migrating (by a factor of 20), with a larger maximum earthquake magnitude. In contrast, activity reported from the Iceland northern volcanic rift zone [White et al., 2011] exhibited both a short-term (∼hours) and long-term (∼weeks) migration. The faster short-term migration at a velocity of ∼4 cm/s is similar to that we observe with the 2009 Mammoth swarm. Both the Tahoe and Iceland swarms have been interpreted as dike injection and propagation [Smith et al., 2004; White et al., 2011].
 While slow, long-lasting migration is probably consistent with magmatic movement (especially with accompanying geodetic deformation), diffusion of lower viscosity CO2–rich fluid might explain the relatively short-duration, fast migration, and limited seismic moment observed beneath Mammoth. CO2 at lower-crustal temperatures and pressures would exist as a supercritical fluid, and with its low solubility relative to H2O, is thought likely to be released in the deep crust from saturated mafic melts [Lowenstern, 2001]. In particular, CO2 would have greater buoyancy and lower viscosity than magma (see NIST chemistry WebBook, webbook.nist.gov), and thus would be expected to move more quickly than magma along zones of weakness in the lower crust. Further supporting this hypothesis is the abundant CO2 release at the surface in the Mammoth area, presumably originating from the asthenosphere [Sorey et al., 1998]. Diffusion of pressurized CO2 has been proposed to explain earthquake swarms elsewhere, including an upper crustal sequence with thousands of events (up to magnitude 6) in northern Italy in 1997 [Miller et al., 2004]. Miller et al.  hypothesized that initial events in the sequence released pressurized fluid, which then diffused outward and triggered the unusually energetic aftershock sequence.
 Diversity in source mechanism in the Mammoth swarms (Figure S2 of the auxiliary material) is a characteristic shared with both the Tahoe [von Seggern et al., 2008] and the northern volcanic rift zone in Iceland [White et al., 2011] swarms. The “flipping” of mechanisms was particularly well documented by White et al. , whose dense temporary seismic deployment showed that both normal and reverse events occurred very close in time, apparently on the same structure as illuminated by the seismicity. The authors offered several possible interpretations of this behavior, including the idea that small jogs between en echelon fingers of the dike might be slipping. This mechanism is an attractive hypothesis to explain the mechanism variation in the Mammoth swarms, since any pressurized fluid could trigger this style of slip.
 The lower crustal earthquake swarms investigated here give clues to the dynamics of Long Valley and Mammoth mountain unrest, including the processes that may underlie shallower seismicity and past eruptive activity. We suggest that the deep brittle-failure earthquakes are occurring within the more mafic lower crust, which can remain brittle to temperatures as high as ∼700o C [Sibson, 1982], especially in the presence of high-pressure fluid [Reyners et al., 2007] and possible sudden stress changes. The plateau in upward migration of the 2009 swarm at ∼19.5 km may reflect a low-permeability barrier associated with transition to more silica-rich rocks and plastic deformation, consistent with observed transition from brittle-failure to long-period earthquakes. In the upper crust (above 6–8 km) temperatures below 350–400°C rocks are again in a brittle regime and capable brittle shear failure.
 The speed, duration, and modest seismic moment suggest that upward migration of CO2 may be a likely trigger for the deep 2009 Mammoth swarm, especially considering its abundant release at the surface. Additionally, the possibility exists that pulses of CO2 may also influence shallow seismicity, and that CO2 might be mobilized upward from the lower crust during deep swarms. While it's unclear over what timescale CO2 propagates through the crust, we note that elevated rates of shallow earthquakes were observed around Mammoth in the months following the 2006 and 2009 deep swarms. Regardless of this timescale, seismicity extending from base of the crust to the surface beneath Mammoth Mountain suggests that this may be a preferred pathway for magmatic volitiles such as CO2-rich hydrous fluids, and occasionally magma, to travel to the surface.
 We are grateful to G. Waite, S. Moran, S. Prejean, and an anonymous reviewer for reviewing this manuscript and to A. M. Pitt, M. Mangan, B. Chouet, W. Evans, and J. Lowenstern for helpful discussions. Waveform data and phase picks were obtained from the Northern California Earthquake Data Center (NCEDC).
 The Editor wishes to thank two anonymous reviewers for their assistance evaluating this manuscript.