Pn travel times are useful for studying crustal and uppermost mantle structure and regional tectonics because they are affected by crustal velocity and thickness as well as uppermost mantle velocity and anisotropy. We obtained 57,740 Pn travel time picks from 5433 earthquakes and 307 stations from Chinese national and provincial earthquake bulletins and the International Seismological Center bulletins to invert for Pn velocity variation and anisotropy and station delays in China. Our inversion reveals significant features that correlate with surface geology. The main results are as follows: (1) The Pn velocities show a mosaic of very fast and very slow anomalies, mirroring the heterogeneous geology of China at the surface. The Pn velocities are high beneath the major basins in the west (Sichuan, Qaidam, west Tarim, Tulufan, and Junggar) and low in areas of active volcanoes (Myanmar and western Yunnan) and Quaternary volcanism in northern Tibet, in seismically active areas in north China and Tien Shan, and in the southern part of south China (the Hainan plume). (2) The Pn anisotropy beneath the major basins in the west is generally weak. Strong anisotropy is found beneath high-deformation regions (the Tibetan Plateau, southeastern margin of the Tibetan Plateau, western Tien Shan, and part of north China), suggesting the anisotropy is likely related to recent large-scale tectonic activity. (3) A large area of north China shows prominent low Pn velocity beneath Archean basement with thin crust. Our observations are consistent with rifting, lithospheric thinning, and mantle upwelling in the region. The Pn anisotropy is consistent with a dextral simple shear in the NNE direction in the lithosphere mantle during the last (and ongoing) major deformation period. (4) The Pn velocity in northern Tibet is generally lower than that in the south. Southern Tibet has significant E-W structure. Low-velocity anomalies can be traced from northern Tibet across southwestern Tibet and south central Tibet to near the India plate. Anisotropy is absent beneath much of the Himalaya block, but consistent anisotropy with E-W fast direction is present beneath the Lhasa block and large anisotropy (up to 4%) is observed in low-velocity regions of the northern and western Tibet. Complex station delays in the eastern margin of the plateau suggest that the whole crust may be highly deformed. The anisotropy pattern in the southeastern margin of the Tibetan Plateau suggests a mantle lithospheric deformation similar to the clockwise rotation of material observed at the surface. (5) Crustal thicknesses inferred from our station delays are consistent with previous models, which correlate well with surface topography.
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.
 China is geologically highly heterogeneous, consisting of Precambrian platforms surrounded by accreted continental fragments and fold belts of various ages (Figure 1). The heterogeneity is most striking with the sharp contrast between the Tibetan Plateau in the west with an average elevation of 5000 m related to the India-Eurasia collision about 60 Myr ago and the Archean core of the Sino-Korean and Yangtsz cratons in the east (Figure 1). The country is seismically very active, even in densely populated areas, such as Sichuan, Yunnan provinces in southwestern China and in the Beijing-Tianjing-Tangshan area in northern China. Thus studies of crustal and upper mantle structure beneath China are of major interests over the years in understanding the tectonics of the region and have important implications for earthquake hazards.
 Global tomographic inversions have yielded important results in the region; for example, van der Hilst et al.  and Grand et al.  showed the signature of the subducted slab of the Mesozoic Tethys sea into the deep mantle. Many high-resolution tomographic studies of the region using regional and local travel time or surface dispersion data have been carried out inside China [e.g., Liu et al., 1990; Liu and Jin, 1993; Xu et al., 2000; Xu et al., 2001; Zhu et al., 2002; Lei et al., 2002; Huang et al., 2002; Xu et al., 2002; Huang et al., 2003a] with most published in the Chinese literature. Over the decades, one of the most significant efforts carried out in China was deep seismic sounding (DSS) using active sources to image the crustal structure. More than 36,000 km of DSS profiles have been acquired in China since 1958 [Li and Mooney, 1998]. They conclude that the upper mantle Pn velocities are about 8.0 ± 0.2 km/s beneath China and the average crust thickness becomes progressively thicker from the east to west with the thickness of more than 70 km beneath the Tibetan Plateau.
 In this study, we use the travel times of Pn waves to invert for the velocity and anisotropy distribution in the uppermost mantle and the crustal thickness of China. A Pn wave can be regarded as a head wave traveling below the Moho along the top of the upper mantle (Figure 2). Crustal thickness and Pn velocity are basic parameters of the Earth's interior. Pn arrival times are routinely used for locating regional events, thus accurate Pn velocity models are important for earthquake location. Pn velocity varies with changes in temperature and material composition and Pn anisotropy may indicate the history of mantle deformation [e.g., Bamford, 1977; Hearn, 1996]. Thus Pn velocity and anisotropy has also become an important tool to probe lithospheric structure and dynamics [e.g., Hearn, 1996; Silver, 1996]. Some Pn inversions haven been conducted in local regions of China, such as the Tibetan Plateau [Zhao and Xie, 1993; McNamara et al., 1997], Xinjiang (western China) [Pei et al., 2002], and southwest China [Huang et al., 2003b], and in the whole country [Wang et al., 2002; Sun et al., 2004]. Here we use Pn arrival time data from the Annual Bulletin of Chinese Earthquakes (ABCE) (a national earthquake catalog compiled by the Institute of Geophysics, China Seismological Bureau), two provincial earthquake bulletins, as well as the International Seismological Center (ISC) bulletins. The large collection of data allows us to adopt stricter criteria on data selection to obtain high-quality data as well as good ray coverage. The tomographic inversion provides a survey of the large-scale structure of the Pn velocity, its lateral variation and anisotropy, and the crustal thickness of China. The model reveals significant features that may have implication for the structure and tectonics of the region.
 The tectonics of China and surrounding areas are complex and challenging. We give a brief summary of the tectonic history of China below, following a few past reviews [Zhang et al., 1984; Li, 1998; Wang and Mo, 1995; Burchfiel and Royden, 1991; Yin and Harrison, 2000]. China is principally a part of the Eurasian plate, except for the Himalayas and the Coastal Range of Taiwan, which are margins of the Indian and Philippine Sea plates, respectively. It consists of nuclei of Precambrian cratons and a mosaic of later accreted microcontinents and fold belts. The fold belts and ancient suture zones are characterized by widespread ophiolite, blueshist, and mélange belts. There are three major Precambrian cratons: the Sino-Korea craton, also known as the North China Block (NCB), the Tarim craton, and the Yangtze craton. The combination of the Yangtze craton and the Precambrian Cathaysia Block, which consists of a part in the south China Fold belt and a part underwater to the east, is also called as South China Block (SCB).
 The Tarim-Sino-Korea paleocontinent accreted to the Angaraian paleoplate (with Siberia craton as nucleus) in the north by consuming the oceanic crust between them. The two continents collided along the Junggar-Hegen suture in the Permian (250–280 Ma). Both the NCB and the SCB were parts of the supercontinent Rodinia in the early Neoproterozoic (at 1.0 Ga). The breakup of Rodinia separated the NCB and the SCB from the other former Rodinian continents. The collision between the NCB and SCB started at about the same time as the north China-Angaraian collision in the Permian. Major suturing between the two blocks occurred during the late Middle Triassic to Middle Jurassic along the Sulu and Qinling-Dabie orogens, which led to the formation and exhumation of ultrahigh-pressure rocks in both regions.
 Two major processes controlled the Cenozoic tectonics of East Asia: (1) the collision of the India and Eurasia plates that started at about 40–70 Ma [Molnar et al., 1993; Rowley, 1996; Yin and Harrison, 2000]; and (2) the subduction of the Pacific (and later the Philippine Sea) plate that started during the late Mesozoic. The continued subduction of the Pacific and Philippine Sea oceanic crust led to the opening of the Japan Sea and the formations of the Western Pacific archipelagos.
 The India-Eurasia collision may have controlled most of the large-scale Cenozoic tectonic history of Asia [e.g., Molnar and Tapponnier, 1975; Tapponnier et al., 1982]. The Himalayan-Tibetan orogen was built upon a complex tectonic assemblage of microcontinents, flysch complexes, and island arcs accreted onto the southern margin of Eurasia since the early Paleozoic. The India-Eurasia collision caused the formation of the Himalayas, thickening of the Tibetan crust, and the uplift of the Tibetan Plateau. At least 1400 km of north-south shortening has been absorbed by the orogen since the onset of the collision [Yin and Harrison, 2000].
 During Cenozoic times, extension occurred along much of the eastern margin of Asia, from Java and Summatra in the south to the Sea of Japan in the north. It is thought that the India-Eurasia collision forced the extrusion of continental blocks eastward along large strike-slip faults, causing the extension in the eastern margin of Asia and the opening of the South China Sea [e.g., Molnar and Tapponnier, 1975; Tapponnier et al., 1982]. This, however, is uncertain. Numerical simulation of the collision with Asia as a viscous sheet deformed by a rigid indenter (India) suggests no extrusion beyond the thickened Tibetan Plateau and thus other tectonic features east of the plateau may not be related to the collision [e.g., England and Houseman, 1986; Royden et al., 1997]. Furthermore, it is argued that the extension in the eastern margin is related to the interactions between Eurasia and the oceanic plates to the east [Burchfiel and Royden, 1991; Northrup et al., 1995].
2. Inversion Method
 We follow the basic inversion scheme of Pn waves by Hearn . The Pn travel time is the sum of the travel time in the crust from the source to the Moho and from the Moho to the receiver, and the travel time of the head wave traveling at the top of the upper mantle (Figure 2). We divide the study area into two-dimensional (2-D) cells of 0.5° by 0.5° along the longitudes and latitudes. Thus the Pn travel time residual (the observed travel time minus the predicted travel time for a reference model) can be written
where tij is the travel time from the ith earthquake to the jth station; tieq and tjst are the travel times from the ith earthquake to the Moho and from the Moho to the jth station, respectively. The earthquake delay Δtieq and the station delay Δtjst represent the discrepancy between the actual travel time and the predicted travel time in the crust, which comes from the departure of the crustal velocity or the crustal thickness or both from the reference model; the earthquake delay Δtieq is also affected by errors in earthquake depth. The travel time perturbation in the upper mantle is separated into contribution from the velocity (slowness) perturbation (dijkΔsk) and contribution from anisotropy (dijk cos(2ϕijk)ak + dijk sin (2ϕijk)bk), where dijk is the length of the ray from the ith earthquake to the jth station in the kth cell of the mantle (dijk is 0 if the ray ij does not pass cell k), sk is the slowness (1/velocity) of the kth cell, and the ak and bk are the anisotropy parameters in the kth cell; the ϕijk is the azimuth of the ray from the ith earthquake to the jth station at cell k; and N is the total number of cells. Here we assume the form of transverse anisotropy for Pn waves, which can be described by a sinusoidal 2ϕ azimuthal variation [Backus, 1965]. Note that we use the azimuth of the ray at each cell (ϕijk) instead of the same azimuth (or back azimuth) of the ray for all cells along the ray [Hearn, 1996], because the azimuth (back azimuth) may change substantially along the ray. The magnitude of anisotropy at cell k is given by (ak2 + bk2)1/2 and the direction of fastest wave propagation is given by arctan (bk/ak)/2 from the north.
 The linear system of equations (1) can be written in matrix form:
where d is the data vector whose elements are the observed residuals of rays Δtij, and the number of the elements is equal to the number of rays; the solution vector m comprises all the model parameters we want to solve: earthquake and station delays (Δtieq and Δtjst), and Pn slowness perturbation Δsk and anisotropy parameters ak and bk in every cell; the matrix G is called the data kernel, which relates the data (observed residuals) and the model parameters and depends on the ray coverage. The matrix G is large and sparse (with a large number of elements being zero). Thus our Pn inversion problem is equivalent to solving a large sparse linear sparse equation system.
 Methods used to solve the linear system such as equation (2) are widely discussed by many authors [e.g., Nolet, 1987]. In this study, we use a preconditioned version of the LSQR algorithm [Paige and Saunders, 1982a, 1982b]. A preconditioning matrix P is applied to (2) to obtain an equivalent system:
where G′ = GP, m′ = P−1m. The LSQR iteratively finds the least squares solution for m′ and the final solution m = Pm′. The preconditioning does not change the least squares norm of the solution, but it accelerates the convergence rate [Paige and Saunders, 1982a, 1982b].
 Preconditioning is needed because the elements of matrix G in equation (2) for slowness and anisotropy are the ray lengths in kilometers, which are 2 orders of magnitude of the matrix elements (which are 1) for station and event delays. Any nonsingular matrix P that approximates the inverse to G can be used as a preconditioner. We choose a preconditioning matrix P below, which is similar to the one used by Hearn and Ni . The matrix makes G dimensionless and takes into account ray path density of each station, event, and cell.
where Pc, Pst, and Peq are diagonal matrices with the dimensions equal to three times the number of cells (N3c), the number of the stations (Nst), and the number of earthquakes (Neq), respectively. The diagonal elements for Pc is given by 1/, where Di is the total ray path length for ray i, and di is the length of ray i within the corresponding cell, the sum is over all rays crossing the cell. The diagonal elements for Pst or Peq are given by 1/, where Nr are the number of rays associated with the corresponding station or earthquake. The dimension of matrix P is equal to the total number of inversion parameters, i.e., Nx = N3c + Nst + Neq.
 Because of uneven distribution of ray paths and data errors, the linear system (2) or (3) is generally highly ill-conditioned, which may result in singularities with extremely high or low values in local areas and make the inversion unstable. A common approach in dealing with the problem is to impose additional constraints to regularize the solutions [Van Der Sluis and Van Der Vorst, 1987]. Here we impose the smoothness constraints as described by Lees and Crosson . Because our slowness solution is a discrete version of the continuously varying field, it is desirable to have some constraints for our solutions to have a certain measure of roughness. Smoothed solutions also allow us to concentrate on more coherent large-wavelength structures and ignore local small-scale variations in interpreting inversion results.
 The smoothness constraints are imposed by minimizing the Laplacian (second derivative) of the solutions m′ (instead of m). Thus our LSQR solutions minimize the following damped least squares functional: ∥G′m′ − d∥2 + λ2 ∥Lx∥2. Here ∥v∥ denotes the Euclidean (L2) norm with ∥v∥2 = (vTv)1/2, and the Laplacian operator L is applied to the component x of m′ that corresponds to slowness or either anisotropic coefficient. The parameter λ controls the level of smoothing, which trades off with the misft (error reduction) ∥G′m′ − d∥. As λ increases, the inversion image becomes smoother and the misfit increases. We choose λ = 0.3 in this study.
3. Data and Model Resolution
 Our Pn arrival time picks are from the Chinese national bulletins ABCE (1985–1997), provincial bulletins by the Yunnan Seismlogical Bureau and the Sichuan Seismological Bureau (1980–1997), and the ISC bulletins (1964–1997). Because of the large number of stations used and the diffusive nature of the seismicity in the Chinese continent and the neighboring region, we can achieve very dense ray coverage of the study area (Figure 3), even after imposing strict controls on data quality.
 The following criteria are used to select the data.
 1. The epicentral distance Δ ranges from 2° to 12°. For a very thick crust of 70 km, the Pn starts to appear at 1.6°; it appears at smaller distances for a thinner crust. We also use arrivals that are labeled as P but continue to form a linear trend in the travel time curve up to distances of 12°.
 2. The earthquake depth is limited to be less than 30 km and 40 km east and west of 107°E, respectively, to exclude limited number of earthquakes that may be in the mantle, e. g. under Tibet [Chen and Molnar, 1983; Zhu et al., 1997].
 3. The apparent velocity of rays is limited to be between 7.0 and 9.0 km/s [Zhao and Xie, 1993].
 4. Every station has at least 10 Pn records to ensure more robust constraint on the station delays and every earthquake has at least 5 Pn records.
 5. We use a method described in Appendix A to mitigate the problem of uneven data distribution. The method aims to locate earthquake clusters and then choose the event with the highest number of observations in a cluster and discard all the other events in the same cluster. This procedure dramatically reduces the unevenness in the data distribution but still keeps the overall ray coverage of the study region intact.
 6. After applying the above selection criteria, we use an iterative linear regression algorithm to derive our initial model for the average Pn velocity and average crustal thickness. The data with travel time residuals (relative to the initial model) larger than 6.0 s are discarded.
3.1. Initial Model
 The formula (5) and (6) below are used to calculate the initial model parameters: the average crustal thickness and the average uppermost mantle Pn velocity m. We choose the average crustal velocity (c) of 6.3 km/s based on the study of Li and Mooney  for our reference model.
where t0 and slope are the intercept time and the slope of the linear regression of travel time as a function of distance, respectively. All the travel times are corrected to the surface focus at the corresponding distances for the regression.
 Following our data selection procedures described above, we finally obtain 57,740 rays from 307 stations and 5433 earthquakes. We obtain the best fitting initial model with Pn velocity of 8.0 km/s and crustal thickness of 43.8 km. All the selected stations and earthquakes are shown in Figure 3a, and the data coverage is shown in Figure 3b.
 To examine our model resolution, we perform two tests: (1) a conventional checkerboard test and (2) a test on an input model that resembles the anomalous regions of the actual inversion. In each test, synthetic data are generated using the real rays, which are then inverted using the same inversion procedure and parameters as in the actual inversion.
 Checkerboard tests are widely used although the method has intrinsic limitations [Leveque et al., 1993]. For Pn velocity tests, our velocity perturbation is a sinusoidal pattern of 0.3 km/s in amplitude with respect to a background velocity of 8.0 km/s. For anisotropy tests, input anisotropy is a sinusoidal pattern of 3% in amplitude and the fast direction alternating at N-S and E-W directions. Several synthetic models with different wavelengths are tested. Results show we can resolve a pattern of 3° in half wavelength very well in almost all the study area for Pn velocity (Figure 4a) and for Pn anisotropy (Figure 4b). The resolution in bordering areas becomes poor. Note that the resolution depends both on the spatial and the azimuthal coverage of seismic rays, e.g., even though we have excellent ray coverage across the Taiwan Strait (Figure 3b), the resolution is still poor. The reason is that the rays in this region have similar azimuths (from earthquakes in Taiwan to stations in the mainland coasts).
 The resolution is quantified in a resolution map (Figure 5) using the resolvability as defined by Zelt . The resolvability, R, from a checkerboard test is defined as
where ti and ri are the true and recovered velocity anomalies at cell i inside a given area of M cells. We choose an operating area of 4.5° × 4.5° centered on the cell for which we want to calculate the resolvability. For the 0.5° cell spacing, M = 81. We use resolvability above 0.7 as an indication of a well-recovered checkerboard structure. The best resolved areas (with resolution of 2° × 2°) are the central part of the country, the north China area, and western Tarim and western Tibet.
 The second test we performed uses a synthesized input model that resembles the anomalous regions of the actual inversion (Figure 6, top). The areas with prominent fast and slow anomalies in the actual inversion are assigned a constant velocity perturbation of 0.3 km/s and −0.3 km/s, respectively. The cells with anisotropy amplitude larger than 1% in the actual inversion are assigned a constant anisotropy amplitude of 3% with the same anisotropy direction. All the other cells are assigned the background Pn velocity (8.0 km/s) with no anisotropy. Synthetic travel times of the real rays are calculated for this model and Gaussian noise is added to the synthetic data using a standard deviation that equals the root-mean-squares (RMS) of the residuals after the real inversion (1.33 s). Despite the large noise level added, the inversion recovers very well the structure of the velocity anomalies (Figure 6, bottom). The pattern of the anisotropy is also generally well recovered, but the anisotropy amplitude is significantly reduced after the inversion.
4. Inversion Results
 Our computation is done under flattened, layered Earth (equations (1)–(7)). However, our final values of Pn velocities and crustal thickness are given in the true spherical Earth, as by Zhao and Xie  and McNamara et al. . To correct for sphericity, we use Earth-flattening transformation [e.g., Aki and Richards, 2002]. The actual crustal thickness H = Re(1 − ≈ Hf(1 − Hf/(2Re)), where Hf is the crustal thickness in the flattened Earth and Re is the radius of the Earth, i.e., H is smaller than Hf by about a fraction of Hf/(2Re). The actual Pn velocity is the Pn velocity in the flattened Earth times (Re − H)/Re. For our reference model with crustal thickness of 43.8 km and Pn velocity of 8.0 km/s, the crustal thickness and Pn velocity in the flattened Earth are 44.0 km and 8.06 km/s, respectively.
 Our inversion results are shown in Figure 7 for Pn velocity and anisotropy and in Figure 8 for station delays. The errors of these parameters are obtained by the bootstrap method [e.g., Tichelaar and Ruff, 1989]. The errors in the Pn velocities are almost all less than 0.05 km/s (Figure 9) (except for bordering areas). Figure 10 shows the ratio of the standard error and the anisotropy amplitude for areas with relatively large anisotropy (amplitude larger than 1.0%). The anisotropy amplitudes are almost all larger than 2–3 times the standard errors. The mean of the standard errors of the station delays is 0.28 ± 0.06 s (2σ). For station delays with amplitude larger than 0.3 s, 96% of them have amplitudes larger than the standard errors and 82% of them have amplitudes larger than two standard errors.
Figure 11 shows travel time residual distribution before and after the inversion. The inversion reduces the RMS error of the residuals from 1.85 s to 1.33 s. The reasons for the relatively small variance reduction (48%) are not clear. Possible causes include (1) data errors, including errors in arrival times picks and misindentification of Pn phases; (2) errors in earthquake locations; (3) the trade-off between the smoothed solution and the misfit; and (4) small-scale variation in crustal velocity and thickness that cannot be accounted for by static station and earthquake delay terms.
 Below we describe the general features of our inversion first and discuss later a few areas of particular interest (north China, the Tibetan Plateau, and Sichuan-Yunnan region) in section 5.
4.1. Pn Velocity Variation
 The Pn velocity ranges from less than 7.7 to over 8.3 km/s. Generally, the Pn velocity in western China is higher than in eastern China. Our inversion reveals significant features that correlate with surface geology. Perhaps the overall most striking feature of the Pn velocity is a mosaic of very fast and very slow anomalies across the region, which provides a mirror image of the heterogeneous geology of China as we briefly summarized in the introduction.
 Prominent high Pn velocities (8.1–8.3 km/s) are found in four major basins bordering the Tibetan and/or Tien Shan (western Tarim, Junggar, Tulufan, Qaidam, and Sichuan basins). These basins are tectonically stable with weak deformation, lack of seismic, thermal, and volcanic activities. Thus the high Pn velocity anomaly is likely the signature that these basins are strong and cold. Note also, not all of the Tarim basin has fast Pn: the values (8.0–8.1 km/s) for the eastern Tarim is only about or slightly above the average.
 Prominent low Pn velocities are observed in several localities: beneath the northern Tibetan Plateau, Tien Shan, the southern part of south China (including northern Hainan), and north China. Pn velocity changes from high (8.0–8.2 km/s) in southern Tibet to very low (7.8–8.0 km/s) in northern Tibet. The low Pn velocity in the north correlates with Quaternary volcanism found in the region [Deng, 1978; Arnaud et al., 1992; Turner et al., 1993]. The result is consistent with previous Pn studies of the Tibetan Plateau [Zhao and Xie, 1993; McNamara et al., 1997]. However, the Pn variation in our model is more complex. The low-velocity anomalies can be traced all the way across southwestern Tibet and south central Tibet to near the India plate along 82° and 90° longitudes, respectively.
 A large area of pronounced low Pn velocity is found beneath the NCB with Precambrian basement. This is in sharp contrast to high velocity generally found beneath old cratons and stable blocks, such as the Canadian Shield and the four major basins in west China. Pronounced low Pn velocity is also observed in south China, beneath much of the Guangxi Autonomous region and the Guangdong province down to northern Hainan. This is consistent with a Hainan plume beneath the hot spot-type volcanism in the region proposed by Lebedev and Nolet . They observed slow S anomaly beneath the same region from the surface down to depths of at least 660 km. The western Tien Shan (west of the Chinese border) shows pronounced low Pn velocity (about 7.8–7.9 km/s), but the eastern Tien Shan (between Tarim and Junggar basins) shows relatively high velocity (about 8.1–8.2 km/s).
 A high-velocity anomaly and a low-velocity anomaly are found beneath Bangladesh and eastern India and beneath Myanmar, respectively. Active volcanoes of the region (Popa, Lower Chindwin, and Singu Plateau volcanoes in Myanmar, and Tengchong volcano in western Yunnan, China) are associated with low Pn velocity. The boundary of the fast and slow anomalies coincides nicely with the India-Eurasia collision zone along the mountain ranges bordering eastern India and Myanmar, suggesting the fast and slow anomalies may be due to strong India lithosphere and weak Eurasia lithosphere as well as the continent-continent collision.
 We also observed slow anomaly beneath the East China Sea margin, the Okinawa trough, and the Taiwan island and fast anomaly beneath Ryukyu trench, which may be related to the subduction of the Philippine plate beneath Eurasia.
4.2. Pn Anisotropy
 The dominant source of anisotropy in continental upper mantle is likely from the lattice-preferred orientation (LPO) of olivine through deformation [Nicolas and Christensen, 1987]. Natural samples of peridotites suggest that olivine a axis (fast direction) is within the foliation plane parallel to the lineation direction [Silver, 1996]. Deformation tends to align the a axis in the direction of major strain axis [Ribe, 1992]. Thus, if the crust and subcontinental mantle deform coherently, the so-called vertically coherent deformation (VCD) model by Silver , then the mantle anisotropy would reflect the last significant deformation environment, e.g., parallel to the direction of maximum shearing under simple shear regime, or perpendicular to the direction of maximum compression under collisional regime, or parallel to the direction of extension under extension or rifting regime.
Figure 7 shows Pn anisotropy with magnitude larger than 1%. There appear some correlations between anisotropy and tectonic activities. Anisotropy in the anomalously fast regions of major basins (Sichuan, Tarim, and Junggar) tends to be small, suggesting little deformation in these stable regions. High values of anisotropy are generally beneath high-deformation regions, including the Tibetan Plateau, western Tien Shan, western Sichuan and Yunnan and Myanmar, and part of north China, suggesting the anisotropy is likely related to present-day mantle deformation. An exception is the region in south central China east of Sichuan Basin and in Guangxi and Guangdong, which is relatively stable with low seismicity, but has strong anisotropy.
4.3. Station Delays and Crustal Thickness
 The station delays can be caused by crustal velocity (Vc) and crustal thickness (H) variations, relative to the reference model (c, ). The station delay is
where η = (1/Vc2 − 1/Vm2)1/2 is vertical slowness in the crust. In most parts of China, the average crust velocities yielded from 36,000 km deep seismic sounding (DSS) range from 6.15 to 6.45 km/s [Li and Mooney, 1998]. Their average velocity for China is about 6.3 km/s. From equation (7), a 0.2 km/s variation in crustal velocity results in about 0.36 s, (assuming = 44 km and m = 8.0 km/s). Thus the observed station delays, ranging from −1.5 to 2.6 s, are mainly from the variation in crustal thickness. If we attribute all the station delays to variation in crustal thickness, one second delay corresponds to 10.2 km difference in crustal thickness from equation (7). Because earthquake depths are often poorly constrained, earthquake delays are strongly influenced by the uncertainty in earthquake depths. For this reason, we do not use the earthquake delay times to infer crustal thickness.
Figure 12 shows crustal thickness contours based on the station delays (Figure 8), assuming the crustal velocities from the global crustal model CRUST 2.0, which is specified on a 2° × 2° cell [Bassin et al., 2000]. The errors in the stations delays, 0.28 ± 0.06 (2σ), would result errors of crustal thickness estimates of about 2.2–3.5 km. If the actual crustal velocity differs by 0.2 km/s from CRUST 2.0, the estimated crustal thickness would have errors of 3 km and 6 km for a 30 km and 70 km thick crust, respectively. Despite these uncertainties, our estimates of crustal thickness from the station delays give an independent, first-order estimate of the crustal thickness variation in China. Because station delays are affected by crustal thickness immediately beneath the stations, they also provide new data coverage for estimating crustal thickness where, for example, DSS profiles are not available. Note also the contours are likely to have large errors in the areas where there are few stations, such as the Tibetan Plateau and all the major basins in the west (Sichuan, Qaidam, Tarim, Junggar, and Tulufan basins). These contours are controlled by stations at the rims of the blocks. Some general features of station delays and crustal thickness are as follows:
 1. With the exception of two stations in the edges of our map (one in Siberia and one in India) (Figure 8), the station delays range from −1.5 to 2.6 s, which correspond to crustal thickness of about 29 to 70 km. The delays in the eastern China are almost all negative, indicating a crust thinner than the average (43.8 km). On the other hand, the delays of the stations in the west are almost all positive, indicating a crust thicker than 43.8 km. Compared with the topography of China, it is clear the difference in crustal thickness between the eastern and western parts of China is due to airy isostasy to the first order.
 2. We observe a sharp change in station delays across the India-Himalaya collision. Almost all of stations on the Indian plate show negative or near zero delays but all the neighboring stations in the Himalayas show large positive delays. This suggests a very sharp change in crust thickness across the boundary (by 10–20 km over a small distance range).
 3. The station delays in the Longitudinal Valley and the Coastal Range in eastern Taiwan are consistently negative while the station delays at the Coastal Plain and West Foothills are either positive or close to zero, suggesting thinner crust along the eastern coast of the island than along the western plain. Note that the resolution for Taiwan is not very good because of poor azimuthal coverage of the rays. Thus we believe the relative station delays are more robust than the absolute delays.
 4. The station delays in the eastern margin of the Tibetan Plateau vary significantly, suggesting large variation in crustal thickness of the region.
5.1. Inversion Without Anisotropy
 To compare with other Pn studies that did not include anisotropy and to examine the effect of including the anisotropy in our inversion, we also conducted an inversion that does not include the anisotropic terms in equation (1) but otherwise follows the same inversion procedure with the same parameters. The inversion results (Figure 13) are generally consistent with those of the inversion with anisotropy (Figure 7). The velocity anomaly patterns are almost exactly the same in the eastern half of the country (east of 100°E), where the ray coverage is considerably better (Figure 3b).
 However, there are some differences in the western half of the country. The fast anomalies near the major basins in the northwest (Tarim, Junggar, Tulufan, and Qaidam) in Figure 13 (inversion without anisotropy) are more scattered in the inversion without anisotropy, while the fast anomalies in Figure 7 (inversion with anisotropy) delineate the location of the basins better. Another important difference is in the Tibetan Plateau. The slow anomalies in Figure 13 are mostly concentrated in the northern Tibet, while the slow anomalies in Figure 7 extend to near the Indian Plate along the longitudes of 82°E and 90°E, respectively. Although our ray coverage in central Tibet is the poorest, our resolution test suggests we can resolve this more complex structure (Figure 6). The results suggest possible trade-off between the velocity anomalies and anisotropy when ray coverage is not sufficient. Thus it is important that the model of the Tibet be further tested with increasing spatial and azimuthal ray coverages.
5.2. Comparison With Previous Studies
 Inversions of Pn travel times in China have recently been conducted by Wang et al.  and Sun et al. . Both studies used a subset of the data we used here. The data used by Wang et al.  and Sun et al.  are arrival times from the ABCE (1986–1996) and the ABCE (1990–1998), respectively. The inversion procedure used by Wang et al. , which also includes Pn anisotropy, is very similar to the one used here; both follow the same basic scheme of Hearn . Sun et al.  obtained a 3-D P velocity model of China using a new tomography method which pieces together individually derived 1-D models (no anisotropy is considered). Their results include maps of Pn velocity and crustal thickness.
 Our results show good agreement with the results of these two studies on large-scale structures. The model of Wang et al.  shows prominent fast anomalies in Sichuan, Qaidam, Tarim, and Junggar basins as in our model. Their model also show the broad slow anomaly in north China and the slow anomaly in Guangxi-Guangdong-Hainan. The patterns of the Pn anisotropy are also generally consistent with our model. The model of Sun et al.  shows clear fast Pn anomalies in Sichuan and Tarim basins, slow anomaly in north China, and relatively slow anomaly in the northern Tibet. It also shows some evidence for fast anomaly in Junggar and slow anomaly in Guangdong.
 However, our model shows considerably better resolution, particularly in the regions of Tibet, the eastern Himalayan syntaxis, and north China. In addition, Sun et al.  show a large area of fast anomalies in the eastern and southern Tibet, which is not evident in our and Wang et al.'s  models.
 Our inferred crustal thickness is generally consistent with previous studies (Table 1). The values for north China are very similar. All models show that the crust thickens rapidly increase across the “N-S belt” of large topography change and active seismicity (around 110–100°E) from about 40 km in the east to about 60 km in the west and that the crust in the Tibetan Plateau thickens toward the central part of the plateau. However, our crustal thickness in eastern China is significantly thicker than those of Li and Mooney  and Sun et al. , but is compatible with a recent crustal thickness map by J. S. Zhu (Chengdu University of Technology, personal communication, 2003), which is based mostly on a large compilation of DSS profiling data. The models of Li and Mooney  and Sun et al.  indicate that the crust of the entire eastern and southeastern China has nearly uniform thickness (ranging from 30 km to 33 or 35 km). It is unclear to us whether the discrepancy is due to resolution or inconsistency of different data sets.
J. S. Zhu (personal communication, 2003), based mostly on DSS profiling.
5.3. North China
 A large area of northern China is underlain by the Sino-Korean (north China) Craton basement, consisting of Archean and Proterozoic metamorphic and igneous rocks [Zhang et al., 1984; Griffin et al., 1998]. Much of the craton remained stable up to about Triassic times, with well-developed sedimentary strata. However, since the late Mesozoic, the region has been tectonically active, with the development of large rifted sedimentary basins and widespread volcanism [e.g., Gilder et al., 1991; Li et al., 1995; Menzies and Xu, 1998; Ren et al., 2002]. The region has high heat flow of up to 100 mW/m2 [e.g., Ma et al., 1984; Liu, 1987] with the average of about 80 mW/m2 in the north China-Bohai basin area. The region is one of the most active areas of intracontinental earthquakes in the world. Four major (M > 7) destructive earthquakes (Xingtai, Bohai, Haicheng, and Tangshan) occurred during the last major burst of seismicity in a decade between 1966 and 1976 [Ma et al., 1990].
 North China stands out in our model. A large area of north China shows prominent low Pn velocity and the crust there is the thinnest of the whole study area. A direct linear fitting of the Pn data that sample the area of longitudes 110–122°E and latitudes 27–42°N yields the average velocity of 7.86 km/s and the average crust thickness of less than 36.6 km. The thinnest crust is located in the north China basin (Figure 12). The elongated shape of the crustal thickness contours, trending NNE, correlates well with the shape of the NNE trending north China basin and the Songliao basin, with the crust thickening beneath the Daxin'anling and Taihang mountain ranges to the west. The boundary along the Daxin'anling and Taihang mountain ranges between the thicker crust to the west and thinner one to the east roughly follows the major lineament of the Bouguer gravity anomaly [e.g., Teng et al., 1983; Griffin et al., 1998].
 The causes for these anomalous observations and the dynamics of the region are poorly understood. Most observations (basin development, volcanism, high heat flow, low mantle velocity, thin crust) are consistent with the notion that the region has been under extension and rifting since Late Mesozoic [Ren et al., 2002]. Furthermore, the high heat flow, Cenozoic volcanism, and low Pn velocity may suggest hot mantle upwelling. Geophysical data seem to suggest a thin lithosphere at present (50–120 km) beneath the basins [Liu, 1987]. However, mantle xenoliths suggest the craton was underlain by an Archean lithospheric keel of 200 km thick [e.g., Griffin et al., 1998], which was replaced by a new lithosphere after the Paleozoic [Gao et al., 2002]. Thus perhaps the mantle lithosphere may have been thinned by as much as 80–150 km during the rifting period in the Late Mesozoic and Cenozoic, by extensional stretching and by convective removal from mantle upwelling.
 Quantitative models of rifting and mantle upwelling in north China (and the larger East Asia margin area) are lacking. Speculation on the mechanisms [e.g., Liu, 1987; Barry and Kent, 1998; Griffin et al., 1998; Ren et al., 2002] often invokes the extension from a far-field effect of the India-Eurasia collision and the back-arc extension and mantle upwelling related to the subduction of the Pacific plate beneath the Eurasia. Extension along the East Asia margin was interpreted as a pull-apart effect due to the eastward ejection of crustal blocks from the India-Eurasia collision zone [Molnar and Tapponnier, 1975; Tapponnier et al., 1982]. However, Northrup et al.  suggested that the extension may be related to the decrease in convergence rate between the Pacific and Eurasia plates from the Paleocene through middle Miocene.
 One significant observation that must be addressed by a realistic model is that the primary strain in north China at present, at least at the midcrust and lower crust, is not that of pure extension but a dextral simple shear in the NNE direction based on the fault slips of major earthquakes [Chen and Nabelek, 1988]. Chen and Nabelek  proposed that the north China basin has developed as a composite pull-apart basin due to right-lateral slip on strike-slip fault systems since the Eocene. Zeng et al. [1995a] suggested that the shallow crust (to 6–8 km) is dominated by extension and normal faulting, but the deeper crust is dominated by strike-slip faults. The change of stress regimes with depth may be explained by magmatic intrusion from the upper mantle to the midcrust.
 Our result on Pn anisotropy in northern China is quite complex. No significant anisotropy is observed beneath the southern part of north China and beneath Bohaiwan with very low Pn velocity. However, strong Pn anisotropy is observed beneath northern Taihang mountain, Yan mountain, and southern Daxin'anling mountain. The fast direction is roughly NNE to N-S, which seems consistent with a dextral simple shear in the NNE direction in the mantle lithosphere. This result is similar to the Basin and Range in the western United States, where mantle anisotropy does not agree with extension direction at surface (i.e., the VCD model of deformation [Silver, 1996] does not apply in these extensional basins).
5.4. Tibetan Plateau
 Our data coverage in Tibet is much denser than those of previous studies [Zhao and Xie, 1993; McNamara et al., 1997]. The number of Pn rays we used is 28,451 for an area (70–110°E, 25–45°N) smaller than that of McNamara et al. , which is an increase of at least 19 times. Excluding the dense ray coverage at the eastern margin, our number of rays is 11866 in the area (70–100°E, 25–45°N). The coverage is still poor compared with the data coverage in other areas of this study (Figure 3b). Note in particular the lack of stations inside the Tibetan Plateau and the Tarim basin. Caution is required when interpreting the crustal thickness contours in those blocks.
 Our models generally agree with Pn velocity inversions of the Tibetan Plateau of Zhao and Xie  and McNamara et al.  (which did not include anisotropy). The agreement is remarkable between our inversion without anisotropy (Figure 13) and that of McNamara et al. . However, our models present a much sharper image and finer scale structure, e.g., of the low anomaly in the Qiangtang block. The difference is likely due to our much increased ray density, which allows us to resolve smaller-scale structure. Our inversion with Pn anisotropy included (Figure 7) shows a more complex finer scale structure of Pn velocity variation in southern Tibet. The low-velocity anomalies can be traced from northern Tibet across southwestern Tibet and south central Tibet to near the India plate along 82°E and 90°E longitudes, respectively. In light of significant mantle anisotropy present in this region from this study and shear wave splitting data, it is important to keep in mind the potential biases from the trade-off between the velocity anomalies and anisotropy when ray coverage is not sufficient. Our observation of the complex structure in the southern Tibet cannot be explained by existing 2-D models of the Tibetan Plateau and suggests that the lithospheric dynamics under Tibet must be 3-D [Molnar et al., 1993; Dricker and Roecker, 2002].
 The pattern of the Pn anisotropy in the Tibetan Plateau is complex. Limited observations of SKS splitting were made previously. For horizontal lineation of LPO olivine (a axis), both the fast Pn direction and the fast polarization of near-vertically traveling shear waves are parallel to lineation direction. Pn anisotropy is absent beneath much of the Himalaya block, which agrees with SKS splitting data [Sandvol et al., 1997; Chen and Ozalaybey, 1998]. Consistent anisotropy is present beneath the Lhasa block. The fast direction is predominately E-W, parallel to the collision boundary (perpendicular to the direction of maximum compression). This is also consistent with SKS splitting data [Silver, 1996]. Large anisotropy (up to 4%) is observed beneath parts of northern and western Tibet with low Pn velocity and the fast direction is nearly NNE to N-S directions. The fast direction departs from the SKS splitting fast direction, which is E-W to ENE, along the Yadong-Golmud Geotransect (between 90°E and 94°E), which is the only profile where a large number of shear wave splitting measurements are available [McNamara et al., 1994; Silver, 1996; Chen and Ozalaybey, 1998]. Because SKS splitting is accumulated along the whole path in the mantle beneath the station, the difference between the Pn fast direction and that of the shear wave may represent a change of mantle (olivine) crystal alignment along the depth of the mantle lithosphere and the asthenosphere. The change of the Pn anisotropy in our model at about 30°N and again at about 33°N agrees well with the changes in SKS splitting and Bouguer gravity anomalies along the Yadong-Golmud profile, observed by Chen and Ozalaybey . They interpreted the changes by the juxtaposition of Indian lithosphere against the overlying Eurasian lithosphere in which the Eurasian lithosphere terminates at 30°N and the Indian lithospheric mantle extends to 33°N. The change in Pn anisotropy between northern Tibet and southern Tibet (Lhasa and Himalaya blocks) suggests that the dynamic regimes of the two regions may be fundamentally different.
5.5. Sichuan Basin and the Interaction Between the Tibetan Plateau and Yangtze Craton
 The pattern of station delays in the eastern margin of the Tibetan Plateau is very complex (Figure 8), suggesting significant crustal thickness variation in the region. Large anisotropy is observed beneath western Yunnan and Myanmar, suggesting strong lithospheric deformation. We also find the surface location of the Sichuan basin is slightly shifted (by about 100–200 km) to the east relative to the high Pn velocity anomaly (Figure 7). The result is quite robust because of the dense ray coverage of the Sichuan-Yunan region (Figure 3b). At the surface, the east margin of the Tibetan Plateau is well developed with north-south striking faults. Mountain belts and rivers turn sharply from nearly west-east direction to the south and southeast direction. The latest GPS observations of the movement of east Asia clearly shows that the moving direction of materials of the region rapidly turns from the east direction to south and southeast [Wang et al., 2001]. In the northwest margin of the Sichuan basin is the Longmen Shan thrust fault belt, which separates the Tibetan Plateau in the west and Yangtze craton in the east [Korsch et al., 1997]. However, to the east, the Sichuan is much more stable, without any major faults or earthquakes. The Pn anisotropy of the basin area is also very small.
 A group of large magnitude of anisotropy orients nearly north-south in central China; to the east is the fault-bearing Yangtze platform and to west is the stable Sichuan basin (Figure 7). This region possibly marks the east boundary of the Sichuan basin. The positive station delays are very large in east of the Sichuan basin but relative small in east, suggesting a Moho dipping to the west.
 If the material of the Tibetan Plateau is extruded eastward [Tapponnier et al., 1982], the Sichuan basin is clearly an important boundary and the above observations suggest strong interaction between the Tibetan Plateau and the Sichuan basin. Basins can be classified as rift (extensional) basins, which subside due to extensional thinning, followed by thermal cooling subsidence, and flexural basins, which is developed by loading by thrust sheets in a compressional regime [Watson et al., 1987]. It is believed that the basins in western China are flexural and those in the east are extensional [Watson et al., 1987]. Basins in the central China (including Sichuan) are more complex [Korsch et al., 1997], but the Sichuan basin was considered to be a flexural basin by Watson et al. . A flexual basin tends to develop a wedge-shaped basin, thickening toward the thrust front [Watson et al., 1987]. A wedge-shaped Sichuan basin with the wedge thickening toward the Longmen Shan thrust fault is clearly consistent with the change in the crustal thickness in the basin and the shift between the basin location and the high Pn velocity.
 Another scenario, which is highly speculative, is that when the plateau material moves eastward, hitting the relatively stable Sichuan basin, the material is channeled to the south; as a result of the eastward extrusion of the plateau, the upper crust of the Sichuan basin is pushed to the east, which explains the shift of the Pn fast anomaly relative to the surface location of the basin. This is possible, following the idea that the deformation of the plateau in the upper crust is decoupled from that of the lower crust, at least around the plateau margins [Burchfiel and Royden, 1991; Royden et al., 1997]. However, this may be difficult because there is little evidence for significant E-W crustal shortening across the Longmen Shan and its adjacent foreland from geological and GPS data [Burchfiel et al., 1995; Chen et al., 2000].
 The complex pattern of station delays may be an indication that the whole crust (including the Moho) of the eastern margin of the Tibetan Plateau is highly deformed, just as the high deformation of the upper crust and the lower crust at the eastern margin that Royden et al.  showed in a numerical simulation.
 The fast directions of Pn anisotropy beneath the plateau southeastern margin form a clear rotational pattern, from the E-W direction in east central Tibet, to the N-S direction in western Yunnan, to the E-W direction in southern Myanmar, and to the N-S direction in western Myanmar. The observation correlates with the results from geology [Burchfiel and Wang, 2003], GPS measurements [Chen et al., 2000], and numerical modeling [Royden et al., 1997] that suggest the crustal deformation in the southeastern margin of the plateau consists of a clockwise rotation of material around the eastern Himalayan syntaxis. The correlation seems to suggest a mantle lithospheric deformation of the similar style.
 1. The Pn velocities of China are characterized by a mosaic of very fast and very slow anomalies. The Pn velocities are high beneath the major basins in the west (Sichuan, Qaidam, west Tarim, Tulufan, and Junggar). The Pn velocities are low in areas of active volcanoes (Myanmar and western Yunnan) and Quaternary volcanisms in northern Tibet, in seismically active areas in north China and Tien Shan, and in the southern part of south China (the Hainan plume).
 2. The Pn anisotropy beneath most areas of the major basins in the west are less than 1%. Strong anisotropy is found beneath high-deformation regions (the Tibetan Plateau, southeastern margin of the Tibetan Plateau, western Tien Shan, and part of north China), suggesting the anisotropy is likely related to recent mantle deformation and large-scale tectonic activities.
 3. Crustal thickness inferred from our station delays are consistent with previous models, which correlate well with surface topography.
 4. A large area of north China shows prominent low Pn velocities beneath the Archean basement. The crust there is the thinnest of China. The thinnest crust is located in north China and Songliao basins. Our observations are consistent with rifting, lithospheric thinning, and mantle upwelling in the region. The Pn anisotropy is consistent with a dextral simple shear in the NNE direction in the lithosphere mantle during the last (and ongoing) major deformation period.
 5. Our model in Tibet shows low Pn velocity in the north and high velocity in the south in general, consistent with previous studies. However, significant fine structures exist. Southern Tibet has significant E-W structure. Low-velocity anomalies can be traced from northern Tibet across southwestern Tibet and south central Tibet to the India plate. Anisotropy is absent beneath much of the Himalaya block, but consistent anisotropy with E-W fast direction is present beneath the Lhasa block and large anisotropy (up to 4%) is observed in low-velocity regions of northern and western Tibet. Station delays in the eastern margin suggest that the whole crust is likely highly deformed. The anisotropy pattern in the southeastern margin suggests a mantle lithospheric deformation similar to the clockwise rotation of material observed at the surface.
Appendix A:: A Source Selection Method Used in This Study
 To reduce uneven distribution of earthquakes, we adopt the following procedure to select events. The procedure aims to locate earthquake clusters and then choose the event with most number of observations in a cluster and discard all the other events in that cluster. (1) Get one event as the center of the first group. (2) Calculate the distance between the second event and the first group, say d21. If d21 > dmin (where dmin is given, depending on the earthquake density and the desired ray coverage), then the event is set to be the center of a new group (the second group), otherwise this event is grouped into the first group. (3) Calculate the distances between the ith event and the existing N groups; if the shortest distance is less than dmin, then add this event to the group corresponding to the shortest distance with this event; otherwise the ith event is set to be the center of the (N + 1)th group. (4) Go through all events by applying step 3. Finally all events are grouped into M groups. (5) Average the epicenters of all events in the same group, and use the averaged locations as the center of the group. Regroup all events so that the distance between the event and its corresponding group center is less than dmin. For an event belonging to several groups, group it into the nearest group. As a result, some groups may be cancelled and some new groups will be added. (6) Repeat steps 4 and 5 until no groups are cancelled and no new group are added. Finally, we choose the event with the largest number of travel time picks in each group for our inversion. All the other events are discarded. The dmin used in this study is 8 km.
 We thank Walter Mooney, the Associate Editor, and an anonymous reviewer for constructive comments, which greatly improved the manuscript. We benefited from discussion with Wang-Ping Chen, Jonathan Lees, Rongsheng Zeng, Jieshou Zhu, Hongzheng Wang, Sitian Li, and Jianye Ren on early version of the manuscript. We thank Youshun Sun for a preprint of their paper. This work was partially supported by National Key Basic Research Program of China (G19998040702) and a William and Flora Hewlett International Research Grant.