We construct fine-scale three-dimensional P and S wave velocity structures beneath the Kanto region, central Japan, by seismic tomography with a spatial correlation of velocities. The Philippine Sea and Pacific plates subduct beneath the Eurasian plate in this area and are imaged with high velocities. Oceanic crust at the uppermost part of these subducting plates is a low-velocity layer. Low-velocity oceanic crust of the Philippine Sea plate subducts to a depth of approximately 80 km. There are also two low-velocity bodies with relatively high VP/VS ratios of 1.80–1.90 in the mantle wedge above the oceanic crust of the Philippine Sea plate. We speculate that the westernmost of these two low-velocity bodies consists of 20% partially serpentinized peridotite, continuous with gabbro in the oceanic crust of the uppermost Philippine Sea plate, while the eastern body is composed of 30% partially serpentinized peridotite. We trace the subducting oceanic crust of the Pacific plate to a depth of ∼120 km. The estimated VP/VS ratio of this layer is 1.85–1.90, which indicates a low probability of molten rock; the gabbroic oceanic crust may have been metamorphosed to garnet-granulite.
 The Kanto Plain in the central Honshu arc is the largest sedimentary basin in Japan, approximately 17,200 km2 in size. The vast Tokyo metropolitan area is located upon this plain, which lies near the triple-trench junction of the Pacific (PAC), Philippine Sea (PHS), and Eurasian (EUR) plates (Figure 1). The PHS plate subducts beneath the EUR plate from the south, while the PAC plate subducts beneath the EUR and PHS plates from the east [e.g., Ishida, 1992]. Severe earthquakes, such as the 1923 Kanto Earthquake (Ms = 8.2), have repeatedly occurred on the boundary between the PHS and EUR plates [e.g., Wald and Somerville, 1995]; thus it is important to obtain a better understanding of the configuration of these plates in order to evaluate their seismic potential and mitigate future earthquake disasters in the Tokyo metropolitan area. In this study, we determine the velocities of P and S waves (VP, VS) in the triple-trench junction in order to investigate the configuration of the three plates beneath the Kanto region.
 The PAC plate subducts toward the west beneath the northeastern Japan arc, marked by a dipping zone of seismic activity. On the basis of seismic tomography, Ohmi and Hurukawa  proposed the existence of low-VP oceanic crust beneath the Kanto region at depths of 50–60 km in a north-south direction at approximately 140.3°E. The depths of the S-to-P conversion interface and S wave reflector at the top of the descending PAC plate have been estimated at ∼100 km [Ohmi and Hori, 2000; Obara and Sato, 1988; Obara, 1989], indicating that a seismic reflector exists on the uppermost surface of the PAC plate.
 The PHS plate also subducts beneath the Kanto district, although from the south. Various models of plate configurations for this area have been proposed using analyses of hypocentral distributions and/or focal mechanisms [Noguchi, 1985; Kasahara, 1985; Ishida, 1992]. The high-velocity (high-V) PHS plate has been imaged by seismic tomography [Ishida and Hasemi, 1988; Ohmi and Hurukawa, 1996; Sekiguchi, 2001]. Ohmi and Hurukawa  found a low-velocity (low-V) zone extending in an east-west direction at depths of 30–60 km, and concluded that this zone represents oceanic crust at the uppermost PHS plate. Kamiya and Kobayashi  proposed the existence of serpentinized mantle wedge with a high VP/VS ratio at a depth of approximately 40 km, above the PHS plate.
 The National Research Institute for Earth Science and Disaster Prevention (NIED) has commissioned two seismographic networks in Japan. The Kanto-Tokai microearthquake observation network (KT-net) in the central Japan region has been in operation since 1979 [Hamada et al., 1982], while the high-sensitivity seismograph network of Japan (Hi-net) has been in operation since 2000 [Obara et al., 2005; Okada et al., 2004]. The purpose of this study is to determine the fine-scale structure beneath the Kanto district using KT-net and Hi-net data in order to detect the low-V layer of oceanic crust at the uppermost PAC and PHS plates.
 The target region covers an area between 34° and 37°N and 138.6° and 141°E, with a depth range of 0–200 km. The average station interval over this area is ∼20 km. Earthquakes analyzed in this study were selected on the basis of two criteria: (1) that selected earthquakes are uniformly distributed over the region, and (2) for earthquakes that occur at the same location, we choose the earthquake with the maximum amount of arrival time data. In all, we analyzed 422,799 P wave and 369,596 S wave arrival times for 15,214 earthquakes recorded at 129 Hi-net and KT-net stations. Figure 2 shows the distribution of seismic stations and selected earthquakes across the study region.
3. Travel Time Inversion
Matsubara et al.  introduced spatial velocity correlation into the method of seismic tomography [Zhao et al., 1992] to determine velocity structure. The spatial velocity correlation is so-called “smoothing.” The correction terms for hypocentral location and P and S wave slowness model parameters are estimated iteratively by minimizing the residuals of the travel times. A three-dimensional (3-D) grid net is used to determine the velocity structure. In the improved method [Matsubara et al., 2004], when calculating the velocity at a point, the slowness is interpolated with that of the surrounding grid nodes. This is because the slowness exhibits a linear relationship with the travel times. The inversion is iterated until the change in the root-mean-square (RMS) of travel time residuals converges. A pseudobending method [Koketsu and Sekine, 1998] is used for seismic 3-D ray tracing. We invert for slowness models and source parameters mutually in every iteration process.
 A conjugate gradient (CG) type solver, the LSQR algorithm of Nolet , is used in the inversion to solve the large sparse system of observation equations, using a covariance matrix of the slowness model that has a nondiagonal element. Paige and Saunders  developed the LSQR algorithm, and Nolet  applied this algorithm to seismic tomography, while Zhao et al.  also used this algorithm that allows the damping matrix with only diagonal elements. In the grid-type tomographic method of Zhao et al. , the resolution of the images is equal to that of the grid, however, to avoid spatial aliasing we need to place more than two grid nodes in the smallest resolvable size of heterogeneity considering the Nyquist wavelength. Matsubara et al.  place many grid nodes and used the LSQR algorithm, extended to an arbitrary damping matrix by Nolet , to introduce the smoothness elements to the damping matrix in order to stabilize the solution. Matsubara et al.  considered not only the damping matrix but also the covariance matrix of travel time data and that of slowness parameters and smoothness. The diagonal elements of the covariance matrix of travel time data are the square of the standard deviation of the picking error, and all the off-diagonal elements are zeros. In addition to the covariance matrix of slowness parameters, the smoothness matrix is considered to contain a positive correlation between adjacent grid nodes, as in Laplacian correlation. The smoothness matrix represents the spatial velocity correlation.
 In the present study we placed grid nodes with large interval spacings at depths of 0–7.5 km, since seismic tomography is unable to resolve shallow structures such as sedimentary layers due to the near-vertical incidence of ray paths arriving at the seismic stations. We calculated the average travel time residual for each station and, in order to remove the effects of local surface structure at each seismic station, added this average to each station correction in every iteration process. If we ignore the station correction, it is possible that the resulting model will contain surface low-V structure at erroneous depths.
 The 1-D structure routinely used to determine earthquake hypocenters at NIED [Ukawa et al., 1984] is employed as the initial model (Figure 3). We do not assume any velocity boundaries such as the Moho discontinuity or the upper boundary of the PAC plate. The interval of the grid nodes and corresponding resolutions are shown in Table 1.
Table 1. Intervals Between Grid Nodes and Checkerboard Patterna
The checkerboard pattern relates to twice an interval of grid nodes.
 The travel time inversion algorithm was used to solve the hypocentral parameters for all earthquakes; 49,650 P wave and 46,550 S wave slowness parameters were used in this process. We solved the VP and VS structures and the hypocenters until the reductions of the RMS of residuals converged.
 The travel time inversion reduces the RMS of the P wave travel time residual from 0.333 to 0.195 s (59%) in 10 iterations. The corresponding RMS residual for the S wave data decreases from 0.667 to 0.302 s (45%).
3.2. Resolution Test
Figure 4 shows the results of a checkerboard resolution test. We assumed a ±5% checkerboard pattern (Table 1) and calculated synthetic travel times using the fast 3-D ray tracing [Koketsu and Sekine, 1998] with random noise of zero mean and standard deviations of 0.05 s and 0.1 s for P and S waves, respectively, derived from the estimated uncertainty of arrival time. The slowness model is then inverted with fixed source. We find that the northern and western parts of the study region are well resolved at a depth of 5 km (Figures 4a and 4b). Almost the entire region beneath the Kanto Plain is well resolved at a depth of 30 km (Figures 4c and 4d). Resolution beneath the southeastern part of the study region at a depth of 5 km is poor due to a lack of seismic stations (Figures 4a and 4b). Figures 4e–4g show a vertical cross section of an area well resolved with respect to S waves. Areas that are well resolved for P waves are almost the same as those areas well resolved for S waves. The lower limit of the area of good resolution is dictated by the distribution of earthquakes. We can resolve to a depth of 150 km (Figure 4e) in the western part of the study region at 139.4°E, and 120 km (Figure 4f) in the central study area at 139.9°E. At 35.8°N, we can only resolve the zone above the PAC plate (Figure 4g), as there are no earthquakes recorded beneath the plate.
 We also conducted a restoring resolution test [Zhao et al., 1992] on the obtained structure including the low-V body. The methods of the ray tracing and inversion are the same as in the checkerboard resolution test except for the hypocentral parameter; in the restoring resolution test, the source parameters are also inverted. Figures 4h–4j show the results of the test; the low-V zone is clearly imaged.
3.3. Station Correction
 The distribution of station correction values for P and S waves is shown in Figure 5. Station correction values for mountainous areas are negative, while those for plains areas are positive. In the Tokyo and Sagami Bay areas, the station correction values range from 0.25–0.80 s for P waves and 0.50–1.50 s for S waves. These values are larger than those in other areas, however, the station corrections for some Hi-net stations, such as FUTH, YKHH, and YROH, are negative. The stations with relatively small station corrections are installed at depths greater than 2000 m. This indicates that borehole seismic observatories at greater depths have lower station correction values even in areas where bedrock is covered by a great thickness of sediments.
 We show the results of the velocity perturbation analysis with respect to the 1-D average velocity model shown in Figure 3. Figure 6 shows estimated VP and VS perturbations as well as the VP/VS ratio at a depth of 5 km in the shallow crust beneath the Kanto area. In shallow layers, the spatial resolution is limited by the station distribution, as the incident angle of the ray path is nearly vertical. In accordance with our data, we placed grid nodes at 0.125° intervals (Table 1). The areas of negative (low) velocity perturbations of 5–8% correspond to topographically low areas of the Kinugawa (approximately 139.9°E and 36.5°N) and Tonegawa Lowlands (approximately 139.2°E and 36.3°N; Figure 5); these areas are covered by >1000 m thickness of soft Neogene sedimentary rocks. In contrast, the areas of 2–5% positive (high) velocity perturbations correspond to the Kanto (approximately 139.0°E and 35.8°N), Ashio (approximately 139.4°E and 36.6°N), and Tsukuba-Yamizo Mountains (approximately 140.2°E and 36.4°N; Figure 5), that consist of hard basement rocks of a Mesozoic accretionary complex.
Figure 7 shows the estimated VP and VS perturbations and the VP/VS ratio at a depth of 30 km. This depth is generally considered to be within the lower crust. The low-V zone is located at approximately 35.75°N and has a width of ∼40 km (Figure 7, region A). This low-VP and low-VS zone extends more than 100 km in an E-W direction. The VP/VS ratio is >1.76 in the low-V zone (Figure 7, region B') and >1.80 at 139.4°E and 139.9°E (Figure 7, region C), all at a depth of 30 km.
 The high-V layer of the PHS plate is dipping at 20° at 34.9°N, and reaches 80 km depth at 36.4°N, 139.3–139.9°E (Figures 8 and 9). A low-V layer overlies this high-V layer, and extends to a depth of ∼80 km (Figure 8, region D). A further low-V zone (8–10%), with VP/VS ratio of 1.73–1.85, is recognized at 35.6°–36.1°N, at depths of 30–50 km (Figure 8, region B). At approximately 139.9°E, the high VP/VS zone is poorly resolved at 30–50 km depth (Figure 9, region between C and E), however, the zone is clearly imaged at depths of 50–60 km (Figure 9, region E) and 80–100 km (Figure 9, region F). The deeper high VP/VS zone (region F) represents the oceanic crust of the PAC plate, while the shallower layer (region E) represents heterogeneity in the PHS plate. Active seismicity occurs along the boundary of the EUR and PHS plates, however, earthquakes occur only rarely in the low-V zone at depths of 30–50 km (Figure 8, region B).
 Vertical cross sections of the estimated VP and VS perturbations and VP/VS ratio at 35.75°N are shown in Figure 10. A significant low-V zone with a perturbation of >8% in both P and S waves is situated at 139°–140.75°E. The depth of the low-V zone between 139°–140°E is 30–50 km; this zone has a high VP/VS ratio. The high-V layer within the PAC plate subducts from 70 km depth at 140.4°E to 150 km depth at 139.9°E. Above this high-V layer, a 2–5% low-V zone of 15–25 km thickness extends from a depth of about 60 km at 140.4°E to ∼120 km depth at 139°E. The low-VP layer is shallower than the low-VS layer. The low-V layer above the high-V PAC plate coincides with the location of earthquakes at the plate boundary.
5.1. Shallow Seismicity
 Within the upper crust of the study area, seismicity is active within a low-VP/VS zone (Figures 6c and 8c). Figure 6 shows those earthquakes at depths of 0–10 km used in this study overlain on the VP/VS structure at a depth of 5 km. Almost all the seismic events lie within the area of VP/VS < 1.70. Beneath the Ashio area (about 139.5°E and 36.7°N in Figure 6), seismicity is active in the low-VP/VS zone at depths of ∼8 km; this zone has high-VP and high-VS and is bounded by the low-V zone (Figure 8).
Tanada  compared a 3-D velocity model for the area beneath the Hakone Volcano, which lies in the southwestern part of the study area (about 137°E and 35.2°N in Figure 6), with the temperature structure and hypocentral distribution. Tanada  found that few earthquakes occur in the low-V (high temperature) zone at depths of 5–10 km; most of the seismic events beneath this region occur in the high-V (low temperature) region at depths of 10–15 km. Our results for the area beneath Ashio are consistent with those of Tanada , who concluded that the events in the 10–15 km high-V zone occurred within the brittle-ductile transition zone. Both the Ashio and Hakone regions are located at the margin of our study area. To investigate whether molten rock exists beneath these areas, we need to expand the study area using finer-scale grids and add data from temporal seismic observations.
5.2. Subduction Zone Geometry of the Philippine Sea Plate
 An analysis of guided waves indicates the presence of oceanic crust of the PHS plate beneath the Kanto district at depths of 30–70 km [Hori, 1990]. At 139.4°E, Hori  observed oceanic crust from 35.8°N, at a depth of 40 km, to 36.2°N, at a depth of 70 km (Figure 8); however, due to the absence of later phases from deeper events, it remains unclear as to whether the oceanic crust extends to depths in excess of 70 km. In our results (Figure 8), the oceanic low-V layer can be traced to at least ∼80 km depth, where earthquakes have been recorded (Figure 8, region D). The subduction path of the PHS plate determined by Ishida  is overlain on our data in Figures 8–10. The plate intersects with region D in Figure 8. We consider the upper boundary of the PHS plate to be shallower than Ishida's estimate for 35.4°–36.2°N, since low-V oceanic crust occupies this area and thrust-type seismicity is recorded in region D. This interpretation of a shallow plate boundary is consistent with that obtained by seismic reflection [Sato et al., 2005].
5.3. Low-V Zone Adjacent to the Philippine Sea Plate
 Horizontally averaged VP values at depths of 10–30 km are a little higher than initial model [Ukawa et al., 1984], while values at depths >40 km are almost uniformly low except for depths of ∼120 km. Averaged VS values for depths >10 km are almost all higher than the initial model except for depths of ∼80 km. These averaged velocities are not evenly distributed about values generated from the initial model, but instead are almost all higher than those of the initial model. The maximum and minimum perturbations are >10%. The large perturbations that occur despite the use of a strong damping parameter indicate that the our final models reflect the actual structure and that the influence of the initial model is reduced. From this point we discuss absolute velocity.
 The low-V zone within region A, Figure 7, has an E-W trend and a width of ∼30–40 km at depths of 30–50 km (Figures 7–10). This zone has been previously documented by Ohmi and Hurukawa , Kamiya and Kobayashi , and Sekiguchi . Ohmi and Hurukawa  suggested that the low-VP zone at depths of 30–60 km and 35.9°N, 139.4°–139.9°E (Figure 8, region B) represents low-density crust of the subducting Izu-Bonin arc stacked within the PHS plate. According to our results, the value of the VP/VS ratio in this low-V zone is less than 1.85. Kamiya and Kobayashi  interpreted the low-V zone as a serpentinized mantle wedge above the PHS plate at depths of 30–45 km and longitude 139.4°E; this interpretation was based on VP and VS values, and a high Poisson ratio of 0.337 equivalent to a VP/VS ratio of 2.02 (Figure 8, region B'). The equivalent VP/VS ratio calculated in this study is 1.80–1.90, which is in agreement with the results of Sekiguchi . Laboratory experiments demonstrate that at 1 GPa, pure peridotite has a VP/VS ratio of 1.83, with VP = 7.9 km/s and VS = 4.3 km/s. Twenty percent partially serpentinized peridotite has a higher VP/VS ratio of 1.87, with VP = 7.3 km/s and VS = 3.9 km/s [Christensen, 1972]. The VP and VS values for the low-V and high-VP/VS zone of region B' (Figure 8) determined in the present study are therefore consistent with the values predicted for 20% partially serpentinized peridotite (Figure 11).
 We consider that the low-V zone consists of gabbro in the uppermost part of the PHS plate and partially serpentinized peridotite above the PHS plate. This zone includes parts of the oceanic crust and the overlying mantle wedge. We can clearly distinguish low-V mantle wedge from low-V oceanic crust on the basis of VP/VS, however, it is difficult to detect the upper boundary of the PHS plate on the basis of physical properties such as VP, VS, and VP/VS. The recognition of a high VP/VS ratio zone (Figure 8, region B') by Kamiya and Kobayashi  may not be precise due to the inaccurate configuration of the Moho in their model and the small amount of S wave arrival time data; Kamiya and Kobayashi  assumed horizontal velocity discontinuity at a depth of 32 km, close to the depth of the Moho, and the number of S wave arrival times used in their analysis was only one third of the number of P wave arrival times.
 The low-V zone at approximately 35.8°N, 139.9°E and a depth of 35 km has a high VP/VS of up to 1.85 (Figures 9 and 10, region C), however, VP = 6.9 km/s and VS = 3.7 km/s (Figure 12), which are lower values than those predicted for the velocities of gabbro or pure peridotite [Christensen, 1972, 1996]. As the VP, VS, and VP/VS ratio of serpentine are lower than those of gabbro or peridotite (Figure 12), this low-V zone is considered to include up to 30% serpentinized peridotite [Christensen, 1972].
5.4. Low-V Oceanic Crust of the Pacific Plate
Ohmi and Hurukawa  located the low-V layer of oceanic crust of the PAC plate at depths of 40–70 km at 140.25°E to 139.0°E. We imaged this layer at depths of ∼60–120 km between 142.2°E and 139.0°E (Figure 10). Matsuzawa et al.  used the P-to-S-converted wave beneath the northeastern Japan arc to locate a low-V layer at the top of the subducting PAC slab to depths of about 150 km. The low-V layer of the oceanic crust is generally considered to result from hydration of the rock. Gabbro, which represents the dominant component of oceanic crust [Yorder and Tilley, 1962; Christensen and Salisbury, 1975] metamorphoses to eclogite under the severe P-T conditions at the base of the continental crust. The temperature of the boundary between the PAC and EUR plates at depths of 50–150 km is ∼400°C [Yamasaki and Seno, 2003]. Iwamori  used temperature boundary conditions appropriate for a 130 Ma plate to calculate the distribution of H2O and melt at 33.8°N along the direction of the PAC plate motion. The subducting PHS plate distorts the corner flow induced by the PAC plate and suppresses thermal recovery. The temperature at the interface between the slab and mantle wedge is ∼370°C at a depth of 100 km [Iwamori, 2000]. The oceanic crust does not metamorphose from gabbro to eclogite under these conditions because this conversion requires a temperature of ∼450°C [Hacker et al., 2003] and pressure of 3.4 GPa (at 100 km depth) [Iwamori, 2000]. The fact that earthquakes occur within this low-V zone argues against the presence of melt. The low temperatures of ∼370°C mean that the oceanic crust exists as garnet-granulite rather than eclogite. Tomographic analysis undertaken in the present study clearly depicts the subduction of low-V oceanic crust to depths of >100 km (Figure 10), as previously simulated by Iwamori .
Obara  identified the S wave reflector at depths of 80–120 km beneath the Kanto district. The position of the reflector coincides with the upper layer of the double seismic zone of the PAC plate. The low-VS zone corresponds to this position and is imaged more clearly than the low-VP zone (Figure 10). Ohmi and Hori  identified a conversion interface corresponding to the upper boundary of the PAC plate at depths of 60–110 km at this same latitude. The upper boundary of the high VP/VS layer located in the present study (Figure 10) coincides with the interface identified by Ohmi and Hori .
 The high-density KT-net and Hi-net seismic networks provide P and S wave arrival time data that help to reveal the fine-scale 3-D VP and VS structures beneath the Kanto region. The high-V PHS plate subducts northward and reaches a depth of approximately 80 km beneath the EUR plate. A low-V oceanic layer on the uppermost part of the PHS plate subducts at an angle of 20° from 34.9°N, and reaches 36.2°N, 139.3°–139.9°E at a depth of ∼80 km. Two low-V bodies are also recognized in the mantle wedge above the oceanic layer of the PHS plate: a low-V body at approximately 139.4°E is composed of 20% partially serpentinized peridotite and is continuous with gabbroic low-V oceanic crust; a low-V body at approximately 139.9°E consists of 30% serpentinized peridotite with VP/VS ratios of ∼1.80–1.90.
 The high-V PAC plate subducts from a depth of ∼70 km at 140.4°E to ∼150 km at 139.9°E. Above this high-V layer, a 2–5% low-V zone exists from 140.4°E at ∼60 km depth to 139°E at ∼120 km depth; the zone has a thickness of 15–25 km. The VP/VS ratio of this low-V layer is ∼1.85–1.90, and the existence of molten rock is unlikely given the low VP/VS ratio. We consider that the low-V zone represents gabbroic oceanic crust metamorphosed to garnet-granulite (but not eclogite) at a temperature of ∼370°C. Our tomographic method clearly images the low-V oceanic crust subducting to depths in excess of 100 km, as previously simulated by Iwamori .
 We would like to thank A. Jin, S. Hori, and S. Sekiguchi of NIED for valuable discussions. This study was supported under a project of the Operation of Seismograph Networks for the NIED and the Special Project for Earthquake Disaster Mitigation in Urban Areas of the Ministry of Education, Culture, Sports, Science and Technology of Japan. We used the GMT software [Wessel and Smith, 1991] for creating the figures. We are grateful to V. Levin, an anonymous reviewer, and the associate editor for helpful comments and encouragement.