Reviewing data from regional permanent seismic stations and a 2001–2002 deployment of portable broadband seismic stations along the Tibet-Tarim border, we identified five unusually deep earthquakes located beneath the West Kunlun range. Application of regional centroid moment tensor analysis puts these events at depths of 40–60 km, with overall uncertainty ranging between 5–8 km. Previous investigations of the region indicate a Moho depth of ∼62 km, implying that these earthquakes occurred near the bottom of the crust. Hence we infer that the lowermost crust beneath the range is brittle and seismically active. Two of these earthquakes are located directly beneath the trace of the Altyn-Tagh fault, with a nodal plane nearly parallel to the local strike of the fault, and a left-lateral sense of slip consistent with the long-term displacement across the fault. The two events suggest that the Altyn-Tagh fault extends through the entire crust in the West Kunlun.
 The strength of the lower crust, and hence the role it plays in the dynamics of continental lithosphere, remains a controversial subject [e.g., Burov and Watts, 2006; Chen and Molnar, 1983; Jackson, 2002; Jackson et al., 2008]. End members of the debate on this topic have colloquially been termed the “jelly sandwich” and “crème brulee” models. While the former argues for a weak, ductily deforming lower crust between relatively stronger upper crust and the upper mantle, the latter posits a uniformly strong crust above a weaker upper mantle. One of the more compelling pieces of evidence that investigators marshal in promoting any particular model is the depth range of ambient seismicity [e.g., Chen and Molnar, 1983; Maggi et al., 2000], the argument being that earthquakes within the lithosphere are necessarily caused by brittle failure and hence are incompatible with ductile rheology.
 While accurate hypocenters of continental earthquakes are critical in assessing lithospheric strength, focal depths are notoriously difficult to constrain from sparse observations of primary phase arrival times (e.g., P and S phases). However, waveforms of secondary arrivals or surface waves are quite sensitive to both focal depth and mechanism, which may be well constrained by waveform modeling even when data from only a few stations are available. For example, teleseismic centroid moment tensor analysis works well for events with magnitudes greater than 5.5 [e.g., Dziewonski et al., 1981]. Similarly, regional centroid moment tensor (rCMT) analysis can be used for smaller events when data from local and regional networks are available [e.g., Dreger and Helmberger, 1993].
 In this paper we report results of rCMT analysis of seismic data recorded in the West Kunlun range at the northwestern edge of the Tibetan plateau (Figure 1). A key tectonic question in this area is the role the plateau-bounding Altyn-Tagh fault (ATF) plays in the interaction between the highly deformed western Tibetan plateau and the apparently much less deformed Tarim basin to the north. The ATF is a left-lateral strike-slip fault at the surface. Teleseismic imaging [e.g., Wittlinger et al., 2004] suggests that the Tarim lithosphere is subducting beneath Tibet along the ATF. However, gravity modeling has suggested that the Tarim lithosphere underthrusts beneath the Tibetan plateau [e.g., Lyon-Caen and Molnar, 1984], and a broken overriding plate is required to reach isostacy in the West Kunlun range [Jiang et al., 2004]. The ATF therefore appears to act as a kind of plate boundary, but whether and how this fault extends into the lower crust, and the extent to which the lower crust is involved in this deformation, are largely unknown. To shed additional light on the role of the ATF in the larger-scale collision, we applied rCMT analysis to earthquakes occurring near it recorded by both short term [Kao et al., 2001; Roecker et al., 2001] and long term (Global Seismic Network (GSN) and GEOSCOPE) networks in the region.
2. Data and Methods
 All of the seismic stations in the temporary field deployments [Kao et al., 2001; Roecker et al., 2001] and permanent seismic networks (the GSN stations NIL, WMQ, and GEOSCOPE station WUS) were equipped with STS-2 sensors, and recorded at sampling rates of 40 SPS (0.025 s) or 100 SPS (0.01 s). Selection of local earthquakes for rCMT analysis was made by reviewing the catalog of the China Earthquake Administration (CEA) [ http://data.earthquake.cn/index.do]. Seismograms for rCMT analysis were chosen based on visual inspection of waveform quality and the azimuthal coverage of recordings on the focal sphere. We restricted our search to earthquakes with M > 4 to ensure reasonable signal-to-noise ratios. The selected seismograms were bandpass-filtered between 0.02–0.1Hz, which greatly reduces the sensitivity of the waveforms to the wavespeed model without significantly diminishing the depth resolution [Huang, 2007].
 The rCMT algorithm we use is based on that of Kao et al. . Using a technique based on that of Zhu and Rivera , the algorithm calculates the Green's functions for each station from a suite of trial focal depths beneath a fixed epicenter. An average 1D (layered) wavespeed model is assumed, although each station can be associated with its own model. After computing the Green's functions, moment tensor inversion is performed for each trial hypocenter. The quality of the fit is defined using a function Ei, as follows:
where fi(t) and gi(t) are the amplitudes of the i-th observed and synthetic waveforms, respectively, while fi(t)max and gi(t)max are the maximum amplitudes of these waveforms. Ei ranges from 0 (perfect match) to 2 (extreme mismatch). These misfit functions are then combined to form an average waveform misfit (E_ave) defined by:
where N is the number of the components used. The hypocenter and moment tensor that minimizes the average misfit function is considered to be the best estimate.
 One should expect some dependence of the rCMT solution on epicentral location; indeed synthetic tests [Huang, 2007] show that epicentral mislocation on the order of 40–50 km can cause large uncertainties of rCMT solutions. In order to reduce these errors, we first tested published epicenters available for each earthquake and compared the inversion results. When the rCMT inversions with published epicenters resulted in large waveform misfits, we relocated the epicenter using a grid search over phase arrival times and wave azimuths [Uhrhammer et al., 2001].
 An anticipated complexity in this analysis is that our Green's function computation assumes a 1D model, while we expect a strong gradient in wavespeeds across the Tibet-Tarim boundary. The western Tarim basin has a thick (∼5 km) sedimentary layer on top of a ∼50 km thick crust, and northwestern Tibet has thicker crust (60 km or more) with a thinner sedimentary cover [Li et al., 2002]. We account for this heterogeneity by using two different 1D wavespeed models (Table S1 of the auxiliary material), depending on whether the station is located north or south of the Tibet-Tarim boundary.
 We estimated source parameters for several earthquakes, but for purposes of this study we focus on five events of particular interest (Table 1) that occurred at unusual depths (from 43 to 60 km) beneath the West Kunlun range. Because the Moho is reported to be at 60–62 km depth in this area [e.g., Levin et al., 2008; Li et al., 2006], these earthquakes, within their depth uncertainties, appear to have occurred close to the bottom of the crust.
Table 1. Source Parameters for the West Kunlun Rangea
Origin Time (hr:mn:sec)
Epicenter (Lat. / Lon.)
The depths are with respect to sea level. Depth uncertainty is determined based on a 5% increase of the minimum waveform-misfit value.
 We performed a number of trial runs to evaluate uncertainties and test the robustness of these results. The first of these involves improving the epicentral location. For example (Figure 2), event 3 has three published hypocenters with relatively shallow depths of 22 km (CEA) and 33 km (PDE and ISC; meaning unconstrained focal depths), and all published epicenters result in poor rCMT inversions with divergent misfit-vs-depth curves and large misfit values at any depth (Figure 2). After relocation, the new epicenter moves westward, and falls on the trace of the ATF. This relocated epicenter also results in an unambiguous global minimum of waveform misfit at a focal depth of 55 km (Figure 2). While experiments with different models show that the focal depth and mechanism obtained for this event are reasonably insensitive to the details of the model chosen (the mechanism is always strike-slip and the depth changes by less than 5 km), the overall waveform misfit can be reduced by decreasing the crustal thickness to 55 km for stations north of the ATF (TSRN and DUWA), consistent with previous studies suggesting that the Moho depth increases from north to south across the ATF [Li et al., 2002; Wittlinger et al., 2004].
 The inference of a deep crustal source for this event is also corroborated by the relative size of Love and Rayleigh waves. Love waves from this event are significantly stronger than Rayleigh waves, with the largest contrasts at stations TSRN and NIL (Figure 3), and also WUS. Inversion results indicate that these stations are close to the nodal planes of the best-fitting mechanism; therefore strong SH radiation is expected. At the same time, synthetic tests show that focal depths on the order of 50 km are required to generate large amplitude Love waves with relatively small Rayleigh waves. In synthetic tests using layered wavespeed models (Table S1 of the auxiliary material) we find that the Love wave amplitude smoothly decreases with increasing focal depth, and there is no sharp change of amplitude as the focal depth crosses the Moho. Hence, while we can not discriminate between a lowermost crustal and an uppermost mantle earthquake based on surface wave radiation, the small amplitude Rayleigh waves require a deep seismic source (e.g., deeper than 22 km depth reported by CEA). This feature is also seen in the waveforms of other earthquakes that occurred at similar depths (Figures S2–S5).
4. Focal Depths, Mechanisms, and Seismotectonic Implications
 The vertical distribution of the five deep earthquakes, along with other well-located deeper events, is shown in a cross section normal to the trace of the ATF in Figure 4. Intriguingly, the hypocenter of events 1 and 3 (depths of 43 and 55 km) are directly beneath the trace of the ATF. One of the nodal planes of each focal mechanism is aligned with the local strike of the fault trace, and both have slip vectors consistent with the known left-lateral displacement of the ATF. The two earthquakes suggest that the ATF cuts through the crust and hence couples the deformation of the upper and lower crust. While four other earthquakes were previously reported in the uppermost mantle beneath our region, with hypocenters located in a 90–96 km depth range (events 6, 7, 8, and 9 [Chen and Yang, 2004; Fan and Ni, 1989]), their correlation with the Moho position remains a subject of debate [e.g., Priestley et al., 2008]. We note that some estimates of the Moho depth south of the ATF reach 90 km [e.g., Wittlinger et al., 2004], which would place the four events right at the crust-mantle transition. Our results demonstrate that the lower crust beneath the West Kunlun range is seismically active and has a brittle rheology. Still, some of these events may have occurred in the uppermost mantle as well. Although the presence of earthquakes in the uppermost mantle is still debated, the distribution of these well-located deeper earthquakes suggests that seismic activity in our study area is not only in the lower crust but likely across the crust-mantle boundary and forms one single seismogenic layer.
 The focal mechanisms of all earthquakes in the West Kunlun are mainly combinations of normal and strike-slip (Figure 4, top right). Only event 9, reported at 90 km depth [Fan and Ni, 1989], has a thrust-faulting mechanism. P-axes of all these focal mechanisms align in a NNE-SSW direction, plunging from nearly horizontal to vertical. This trend is close to the Tibet-Tarim convergence direction derived from GPS observations [Wang et al., 2001]. Most T-axes are sub-horizontal and aligned in a nearly east–west direction. Subduction of the Tarim lithosphere in this region has been inferred from teleseismic imaging [e.g., Wittlinger et al., 2004]. However, the oblique strike-slip mechanisms of these deeper earthquakes are not consistent with contemporaneous subduction, as down-dip extensional mechanisms should be expected at the depth range of these earthquakes. Hence, our observations do not corroborate the inference of subduction [e.g., Wittlinger et al., 2004]. While most source mechanisms reflect north–south compression and east–west extension, they are more likely associated with a collision between the deeper parts of the Tibetan lithosphere and the Tarim lithosphere.
 We thank Honn Kao of the Canadian Geological Survey and Wen-Tzong Liang of the Institute of Earth Sciences, Academia Sinica, Taiwan, for archiving and releasing the seismic data, as well as the rCMT software. Seismic data used in this study are processed with the SAC software. All maps are generated using the GMT software [Wessel and Smith, 1991]. This study is supported by the National Science Foundation (grant number 431467).