We use an iterative nonlinear travel time tomography technique to determine three-dimensional P-wave and S-wave seismic velocity models for the region beneath the Kanto-Tokai district of central Japan. Salient points of the models are the following: (1) The subducting Philippine Sea slab has thickness variation with a stepwise offset east of the Izu Peninsula. The eastern (beneath the Kanto district) and western (beneath the area north of Izu Peninsula and the Tokai district) regions have respective thicknesses of 60 and 25 km. The slab clearly subducts continuously to a depth of 100 km from the Sagami Trough beneath the Kanto district, although the Tokai slab is indistinct at depths of 50–70 km. (2) The oceanic crust at the top of the descending Philippine Sea slab transforms into eclogite at a depth of about 70 km beneath the Kanto district. (3) Low-velocity anomalies are associated with present and past volcanism. They are located on the west of a 110-km depth contour of the Wadati-Benioff zone, rather than only on the backarc side of the volcanic front. Deflection of the volcanic front might result from the subducting Philippine Sea slab, which shifts the dehydration reaction deeper as it subducts deeper.
 Two oceanic plates subduct beneath the Japanese islands. The Pacific plate subducts westward from the east and the Philippine Sea plate subducts northwestward from the south. The two plates are mutually overlain or colliding beneath the Kanto-Tokai district of central Japan. Because of these subductions, seismicity and tectonics are extremely complex in this region.
 Since seminal studies by Aki and Lee  and Aki et al. , three-dimensional seismic structures have been investigated in various regions on various scales. Particularly, many seismologists have applied tomographic techniques in the Kanto-Tokai district of central Japan (Figure 1), thereby revealing details of seismic velocity structures and facilitating discussion of the descending Philippine Sea slab's configuration. The three-dimensional seismic velocity models beneath this district were estimated by Horie and Aki , Ishida , Ishida and Hasemi , and Ishida  using local earthquakes, with block sizes larger than those used in recent studies. Lees and Ukawa , Ohmi and Hurukawa , and Sekiguchi [2000, 2001] estimated velocity models using small block sizes, whereas Lees and Ukawa  estimated the model only in a smaller region around the Izu Peninsula. Ohmi and Hurukawa  estimated a P-wave velocity model from not so many arrival times. Sekiguchi [2000, 2001] used a data set including data with large errors (less than 1.0 s) and did not perform three-dimensional seismic ray tracing in tomography, which might blur the estimated velocity models. Kamiya and Kobayashi [2000, hereinafter referred to as KK2000] determined P-wave and S-wave velocity models beneath the Kanto-Tokai district. However, the configuration of the Philippine Sea slab was not discussed because the resolutions in the shallow region were poor.
Sekiguchi  estimated a three-dimensional attenuation structure beneath this region using spectral ratios of seismic P and S phases and discussed the tectonics and the temperature distribution of this region. Nakamura et al.  also estimated a three-dimensional attenuation (Qs) structure from strong motion records and found that a low-Qs region coincides with the high Poisson's ratio region estimated by KK2000.
 Various models of the configuration of the descending Philippine Sea slab have been proposed not only from tomographic images but also from hypocenter distribution, focal mechanisms, and later phase analyses [e.g., Shimazaki et al., 1982; Nakamura et al., 1984; Noguchi, 1985, 1996; Kasahara, 1985; Ishida, 1992; Hori, 1997]. Recently, Sato et al.  estimated the depth of the upper surface of the descending Philippine Sea slab using a seismic reflection technique with active source in a depth range shallower than 25 km. Hori  proposed a geometry model of the descending Philippine Sea plate beneath the Kanto region. He delineated the upper and lower surface of the descending Philippine Sea slab from the seismicity, the focal mechanisms, and analyses of the later phases recorded in seismograms, which is considered to include seismic waves conducted through the oceanic crust of the uppermost part of the Philippine Sea plate. Kimura et al.  determined spatial distributions of repeating earthquakes and proposed a new plate configuration model in the eastern part of the Kanto district.
 On the other hand, several studies have indicated a thicker Philippine Sea plate before subduction based on source mechanisms and seismicity in the northern Philippine Sea plate. Seno  pointed out that the Philippine Sea plate can be about 60 km thick in the eastern rim of the plate from source mechanisms of an intraplate bending earthquake. Moriyama et al.  also estimated a similar thickness of the Philippine Sea plate from depths of aftershock events of the 1972 Hachijo-Oki interplate earthquakes between the Pacific and Philippine Sea plates, which were distributed down to a depth of 70 km, suggesting that the rigid Philippine Sea plate is present at that depth.
 In general, volcanic arcs are almost parallel to the trenches in subduction zones. The volcanic front, which is defined as a trenchward boundary of the volcanic arc in subduction zones [Sugimura, 1960], is usually parallel to a strike of the descending oceanic slabs and is located about 110 km above the Wadati-Benioff zone [Tatsumi, 1986]. However, in central Japan, the volcanic front deflects westward toward to the northwestern Kanto region. The maximum depth of the Wadati-Benioff zone of the descending Pacific slab corresponding to the volcanic front is about 160 km. Consequently, a cusp region with no Quaternary volcanoes (35.3°N–36.5°N, 138.2°E–139°E) (Figure 1) is formed, although Tertiary volcanic rocks have been found in the cusp region [Nakamura, 1962].
 The present study obtains three-dimensional P-wave and S-wave seismic velocity models beneath this district by employing an iterative nonlinear tomography technique with three-dimensional ray tracing. We used much more P-wave and S-wave travel time data than those used by KK2000, especially from shallow events, to improve the resolutions in shallow regions. We use only data with small errors (less than 0.3 s) because not only quantity but also data quality is necessary to obtain fine models. The estimated P-wave and S-wave velocity models offer much better resolution than KK2000, especially in the shallow regions beneath the southwestern Kanto region, where seismicity is very active. We discuss the transformation depth of gabbro into eclogite within the oceanic crust, the variation of the thickness of the Philippine Sea slab, low-velocity anomalies associated with volcanism, and its relationship with the Pacific and Philippine Sea slabs.
 In all, we selected 18,805 hypocenters and adopted 413,803 P-wave and 200,575 S-wave arrival time data for 1979–1996 from the catalog of the National Research Institute for Earth Science and Disaster Prevention (NIED). Total data sets are about 1.5 and 2.2 times larger for P-waves and S-waves, respectively, than those of KK2000. Stations used in this study were 106. Figure 2 shows distributions of epicenters and stations that were used in this study.
 Arrival time data picked by NIED are classified into four ranks, A–D, according to the picking accuracy. Respective errors of arrival times for A, B, C, and D are less than 0.1 s, 0.1–0.3 s, 0.3–1.0 s, and greater than 1.0 s [Ukawa et al., 1984]. We use only data that are ranked A or B by NIED for both P-waves and S-waves to determine precise structures in the same manner as that of KK2000. We select earthquakes whose respective arrival time data of P-waves and S-waves are not less than 10 and 5. For this study, we add events shallower than 32 km to those of KK2000 to improve resolutions over those of KK2000, especially in the shallow region. For shallow events, we sometimes observe earthquakes whose hypocenters are concentrated within very narrow areas, e.g., earthquake swarms. For our tomographic study, we should thin out these earthquakes because data from these earthquakes have less independent information among them. Then we divide the studied region into boxes of 0.01° × 0.01° × 1 km in the Earth interior; we take one event in each box applying the following criteria in order of precedence: (1) An event whose S-wave data are most numerous among the events in the box. (2) An event whose P-wave data are most numerous among the events in the box. (3) The latest event among the events in the box. The numbers of additional hypocenters, P-wave, and S-wave arrival time data over those in the KK2000 study are 6316, 130,894, and 110,299, respectively.
 We conduct the same procedure as KK2000 for iterative nonlinear travel time tomography. We use source parameters determined by NIED as initial input parameters and start from the lateral homogeneous velocity models used by NIED for routine hypocentral determination [Ukawa et al., 1984], which is shown in Figure 2. Using this model, we calculate P-wave and S-wave travel time residuals, which are used in the inversion. After correcting travel times for Earth's ellipticity [Dziewonski and Gilbert, 1976] and station elevation, we improve P-wave and S-wave velocity models using an inversion technique. We trace seismic rays for each pair of hypocenters and stations using the obtained lateral heterogeneous velocity models and relocate the hypocenters using least squares method. We then improve P-wave and S-wave velocity models again. We repeat the improvement of velocity models and seismic ray tracing by turns until the model parameters converge sufficiently. For three-dimensional inversion, we use an algebraic reconstruction technique (ART) associated with Bayesian constraints [Herman et al., 1979; Herman, 1980; Hirahara, 1988]. This method enables us to determine both source parameters and slowness perturbations simultaneously using a priori estimation errors for each datum and model parameter. For further details of this method, see Hirahara .
 For the standard deviations for data and model parameters, we adopt the following: σA = 0.1 s (for data ranked A), σB = 0.2 s (for data ranked B), σi = 2.5% (for slowness perturbations), σϕ = σλ = 0.01° (for latitude and longitude corrections of hypocenters), σh = 1.0 km (for depth corrections of hypocenters), σt = 0.1 s (for origin time corrections), and σSC = 0.1 s (for station corrections). For the relaxation parameter, λ = 0.1 is adopted for each iterative step.
 Resolution analysis is very important to judge the reliability of estimated model parameters. The resolution depends on the number and direction of seismic ray paths. A large number and various directions of ray paths in a region make the resolution better, whereas a small number and narrow range of directions of the ray paths cause poor resolution. As in the work of Hirahara , we can obtain resolution matrices using ART-type methods. However, we do not calculate resolution matrices because the calculations are far too numerous and require too much computing time. We therefore performed checkerboard resolution analyses [Inoue et al., 1990] to verify the reliability of the estimated parameters.
 The modeling space includes the ranges of 32.8°–37.3°N, 136.8°–141.3°E, and depths of 0–450 km. All earthquake hypocenters and seismic stations used in this study are located inside the modeling space. We assume three-dimensional grid nets in the modeling space. We calculate the seismic velocity at any point in the Earth's interior using a simple interpolation function [e.g., Thurber, 1983]. Partial derivatives of travel time with respect to velocity perturbations at each grid point and source coordinates are computed using three-dimensional seismic ray tracing in the estimated lateral heterogeneous media at each step. The algorithm developed by Um and Thurber  is adopted for tracing seismic rays. We also use a small grid interval to reveal fine structures, as KK2000 of 0.1° × 0.1° × 8 km for depths of 0–40 km, 0.1° × 0.1° × 10 km for depths of 40–200 km, 0.1° × 0.1° × 25 km for depths of 200–400 km, and 0.1° × 0.1° × 50 km for depths greater than 400 km. The estimated unknown parameters are 18,805 × 4 source parameters, 106 × 2 station corrections, and 31,192 velocity perturbations for P-waves and 28,792 for S-waves. Our results are completely free from any a priori assumption about the configuration of the subducting slabs.
 We present the upper 100 km of seismic velocity models in Figures 3–13 because the resolutions are poor at depths greater than 100 km for both P-wave and S-wave structures. Checkerboard-resolution test results are also shown in Figures 3–13. Travel-time residuals are reduced from 0.432 s to 0.389 s for P-waves and from 0.761 s to 0.678 s for S-waves after inversion.
 In both P-wave and S-wave velocity models, low-velocity anomalies are remarkable in and around the Tokyo Bay area and in the western Suruga Bay area at depths of 0 and 10 km. The low-velocity anomalies correspond to the Quaternary sediment distribution [Geological Survey of Japan, 1992]. Actually, Quaternary sediments should not be located at such depths. However, low-velocities of sediments are mapped at 10-km depth because of the poor resolution at 0-km depth and the interpolation scheme mentioned above.
 High-velocity anomalies appear in and around the Sagami and Suruga Bays at depths of 20 km. The anomalies in and around the Sagami Bay are attributed to northwestward subduction of the Philippine Sea plate because the northern ends of the high-velocity anomalies locate at the upper surface of the descending Philippine Sea slab delineated by Hori .
 Checkerboard resolutions for both P-wave and S-wave velocity structures are very good throughout the modeling space except for its rim. They are much better than those of KK2000 for 8–24 km depths because we increase data from shallow events substantially. Resolutions at 0-km depth are not improved because they are controlled strongly by the station distribution.
 In both P-wave and S-wave velocity models, low-velocity anomalies are found near the volcanic chain on the west of 139°E. However, the anomalies appear not only on the backarc side of the volcanic front but on both sides of the front. The eastern boundary of the low-velocity anomalies lies at a depth contour of 110 km of the Wadati-Benioff zone of the descending Pacific slab. Similar features are also apparent in results of previous studies [Sekiguchi, 2000, 2001; Honda and Nakanishi, 2003].
 Two other low-velocity anomaly zones were detected beneath the Kanto district. One is the east-west oriented zone beneath the Saitama prefecture at 36°N for longitudes of 139°E–141°E in this depth range. The other has north-south orientation from northern Chiba to southern Ibaraki prefecture at 140.3°E for latitudes of 35.5°N–36.2°N at a depth of 50 km. Similar features were found in previous studies [Ohmi and Hurukawa, 1996; Sekiguchi, 2000, 2001; Matsubara et al., 2005]. We consider that these low-velocity anomalies comprise two parts: one is oceanic crust at the top of the Philippine Sea slab and the other is serpentinized peridotite at the bottom of the wedge mantle. This will be discussed below in detail with the north-south vertical profile.
 High-velocity anomalies are apparent in the vicinity of the Tokyo Bay beneath the Kanto district and in the vicinity of the western part of the Shizuoka prefecture beneath the Tokai district. In this depth range the high-velocity anomaly region beneath the Kanto district locates southward from the contour lines of the upper surface of the descending Philippine Sea plate [Hori, 2006]. Low-velocity anomalies appear between this high-velocity region and the contour lines of the upper surface of the oceanic plate. This high-velocity region has a north-south oriented edge beneath the western Kanto region at approximately 139°E. These features are not an artifact caused by the lack of resolution: the checkerboard resolution analysis suggests good resolution in this depth range, especially beneath the Kanto district. Similar features also appeared in the results of Sekiguchi [2000, 2001].
 In this depth range, high-velocity anomalies were detected beneath the Kanto region. The high-velocity region approaches the contour lines of the upper surface of the Philippine Sea plate [Hori, 2006] again. Other high-velocity anomalies appear in the southern part of Nagano prefecture (35.1°–36°N, 137.2°–138.2°E). They are less clear than those of the Kanto district because of the poor resolution. However, the high-velocity region lies next to the contour lines of the upper surface of the descending Philippine Sea plate [Noguchi, 1996]. Consequently, both high-velocity regions beneath the Kanto and Tokai districts correspond to the descending Philippine Sea plate.
Sekiguchi [2000, 2001] obtained similar features of high-velocity anomalies and pointed out that the slab continues from the Kanto district to the Tokai district at a depth of 70–80 km. However, in the present results, it remains unclear whether the slab is separated or not in this depth range because the high-velocity anomalies are obscured immediately beneath the volcanic front.
 Low-velocity anomalies are located near the volcanic front. These anomalies spread on both sides of the volcanic front, as shown in the shallower part mentioned above. The eastern boundary of the low-velocity anomalies is located immediately west of the descending Philippine Sea slab.
Figure 14 shows the north-south vertical profile of estimated P-wave and S-wave velocity models and the Poisson's ratio distribution obtained from the velocity models at 139.4°E. The upper surface of the descending Philippine Sea slab delineated by Hori  is consistent with our velocity models beneath the western Kanto region. The upper surface is located at the top of the high-velocity zones shallower than about 25 km. At depths of 25–70 km, it is located within a low-velocity zone located above the high-velocity region. The boundary of the high-velocity and low-velocity zones beneath the surface runs almost parallel to the surface. The distance between them is about 10 km. At depths of 25–50 km, a high Poisson's ratio (up to 0.35) region appears only on the upper side of the surface. At depths greater than 70 km, the upper surface becomes the upper boundary of the high-velocity anomalies again. These features are common for both P-wave and S-wave velocity models, although they are clearer for the P-wave than the S-wave model.
5.1. Descending Philippine Sea Slab
 The P-wave velocity of the oceanic crust of the descending Philippine Sea slab beneath central Japan is about 6.6–6.8 km/s [Nakanishi et al., 2002], which is higher and lower than the reference model shown in Figure 2 at depths shallower and deeper than 25 km, respectively. However, gabbro, which is the dominant component of the oceanic crust [Yoder and Tilley, 1962; Christensen and Salisbury, 1975], transforms to eclogite under more extensive P-T conditions. Therefore a transformation of gabbro into eclogite within an oceanic crust would take place as the oceanic crust subducts deeper. Because eclogite is indistinguishable from harzburgite in terms of wave speeds [Hacker et al., 2003], the subducted oceanic crust would no longer show a low-velocity anomaly if the transformation occurred. It would show a high-velocity anomaly in the same way as the subducting mantle slab.
Christensen  compiled compressional wave velocity (Vp) and shear wave velocity (Vs) data of numerous crustal and mantle rocks measured in laboratory. That study revealed that dependencies of the ratios of compressional to shear wave velocity (Vp/Vs) and Poisson's ratios on both pressure and temperature are very small. Consequently, they are applicable for placing constraints on rock composition. That study also showed that serpentinized peridotite, which is generated with mantle peridotite and H2O, is characterized by a low wave speed and a high Poisson's ratio. As a result, we can distinguish between the oceanic crust and serpentinized peridotite in low-velocity anomaly regions located near the upper boundary of the subducting oceanic slab using Poisson's ratio.
 In our velocity model, as mentioned above, the upper surface of the descending Philippine Sea slab is situated within the low-velocity anomaly region at depths of 25–70 km. The lower side of the low-velocity region is inferred to correspond to the subducting oceanic crust and the upper side is inferred to correspond to the serpentinized peridotite because of the high Poisson's ratio only on the upper side.
 The thickness of the low-velocity oceanic crust layer is about 10 km. This oceanic crust layer should continue from the shallower region. However, the layer shows a high-velocity anomaly because the seismic velocity in the oceanic crust is higher than the reference model, as mentioned above. That low-velocity anomaly disappears and changes to a high-velocity anomaly at the depth about 70 km in the oceanic crust layer. For that reason, we can recognize that the transformation from gabbro into eclogite takes place at this depth.
Hori  concluded that the oceanic crust remains without transformation into eclogite at depths deeper than 60 km beneath the Kanto district from analyses of the seismic waves conducted through the oceanic crust of the Philippine Sea slab. This result is consistent with our velocity model. Matsubara et al.  found the low-velocity oceanic crust at the top of the descending Philippine Sea slab to a depth of ca. 80 km in the same vertical cross section with this profile, although the deeper part of the low-velocity crust locates above the upper surface of the slab in their velocity models.
5.2. Thickness Change of the Descending Philippine Sea Slab Beneath the Kanto District
 We depict the P-wave and S-wave velocity vertical profiles A-A' in the WSW-ENE direction (see Figure 15) in Figure 16: their patterns of P-wave and S-wave velocity variations are similar. For discussion of the slab thickness, we adopt the upper surface model of the Philippine Sea slab delineated by Hori  and Noguchi  beneath the Kanto and Tokai districts, respectively, because Hori  is very consistent with our velocity model, as described above. As a lower surface of the Philippine Sea slab, we take the bottom of the high-velocity anomalies or the upper surface of the Pacific slab delineated by Ishida . These upper surfaces are represented as solid lines and the bottom of the high-velocity anomalies is represented as dashed lines in Figure 16.
 The upper surface of the Philippine Sea slab delineated by Hori  is situated at the top of the high-velocity anomaly region beneath the Kanto district at depths about 15–20 km in this cross section. The surface corresponds to the upper boundary of the subducting oceanic crust of the Philippine Sea slab because seismic wave speeds of the oceanic crust are higher than those of the island arc crust in this depth range, as discussed above. A lower boundary of the high-velocity region is apparent at depths of about 70–75 km. The thickness of the high-velocity region is about 60 km. We infer that the entire high-velocity region corresponds to the Philippine Sea slab, which includes the oceanic crust and the slab mantle because it is accompanied by background seismicity as a whole. The location of the lower boundary of the slab is well constrained by ray paths from earthquakes of the descending Pacific slab beneath the Philippine Sea slab in the Kanto region. The lower boundary of the high-velocity region jumps from the depth of 70 km to a depth of 35 km at the western Kanto region. The slab is about 25 km thick on the western side, which is almost consistent with results of previous studies estimated from surface wave analyses [e.g., Abe and Kanamori, 1970]. We find that the slab thickness changes abruptly. From this and other several WSW-ENE vertical profiles, we pick western end points of the depth change of the lower boundary of the slab and plot them onto the horizontal map (Figure 17). The plotted points are located about 30 km east of the Quaternary volcanic chain.
 About 60-km-thick inclined high-velocity zones along with seismic planes also appeared in east-west cross sections at 34.75–35°N of the obtained velocity models by Ishida and Hasemi  and Ishida , although those studies indicated the slab thickness as 30 km and 30 ± 5 km, respectively. A 50-km-thick high-velocity zone accompanied with background seismicity beneath the Kanto district also appeared in the results of Nakajima and Hasegawa . Sekiguchi  also described a 50-km-thick high-velocity zone of the descending Philippine Sea slab east of the Izu peninsula and suggested that the slab is thickened. Hori  found that the hypocenters occurring within the subducting Philippine Sea slab are distributed in a region whose thickness is about 50 km beneath the Kanto district from a vertical view of a northwest-southeast profile of the relocated hypocenter: consequently, the descending Philippine Sea slab is 50 km thick or thicker. Hori  delineated the upper and lower surface of the descending Philippine Sea slab from the seismicity and the focal mechanisms. He found that the average distance of the two groups of hypocenters occurring at the upper and lower surface of the slab is approximately 40 km. He suggested that the slab might be thickened in the north-south direction by the pressure of the descending Pacific plate.
 Several explanations can be given for the thicker slab beneath the Kanto region: (1) the thick part of the slab might have been thick before subduction, as pointed out by Seno ; (2) the thick part of the slab might have been thickened in the north-south direction by the pressure of the descending Pacific plate, as suggested by Hori ; (3) the thick part of the slab might have fractured and separated into two pieces and then mutually overlapped. The actual reason remains unknown, but some relationships exist between the thick part of the slab and Tertiary volcanic activities.
5.3. Low-Velocity Anomalies Associated With Volcanism
Iwamori  pointed out that, based on numerical modeling of mantle flow, temperature, and fluid distribution, the descending Philippine Sea plate delays the thermal recovery of the subducting Pacific plate and shifts the dehydration reaction to greater depths along the Pacific plate. Consequently, magmatism occurs above the deeper Wadati-Benioff zone.
Figures 6–13 show that low-velocity anomalies extend to both sides of the volcanic front at depths of 30–100 km rather than on the backarc side of the volcanic front. The low-velocity anomalies are bounded on the east by the high-velocity Philippine Sea slab. The eastern edge of the low-velocity anomalies is located along the 110-km depth contour of the Wadati-Benioff zone of the descending Pacific slab at 40 km depth; it moves westward as the Philippine Sea slab subducts deeper. Broad, low-velocity anomalies that extend across the volcanic front are also apparent in the vertical profile. In WSW-ENE vertical profiles (Figure 16), low-velocity anomalies are apparent on both sides of the volcanic front just west of the descending Philippine Sea slab. Low-velocity anomalies on the eastern side of the volcanic front also appeared in the results of Sekiguchi [2000, 2001] at depths of 35–65 km. Moreover, Tertiary volcanic rocks have been found on the surface between the volcanic front and the 110 km depth contour of Wadati-Benioff zone [Nakamura, 1962].
 From these observations, we infer the following. Tertiary volcanic activities occurred at the 110 km depth contour of Wadati-Benioff zone, as is apparent in other subduction zones. Because the Philippine Sea slab has subducted deeper, the dehydration reaction at the surface of the Pacific slab has been shifted deeper, as pointed out by Iwamori ; furthermore, the volcanic front has been deflected to the present location. The present low-velocity anomalies between the volcanic front and the 110 km depth contour of the Wadati-Benioff zone, which are shown in our tomographic images at depths of 30–100 km, might be remnants of Tertiary volcanic activities.
 Using an iterative nonlinear travel time tomography technique, detailed P-wave and S-wave velocity models were determined with unprecedented accuracy and resolution beneath the Kanto-Tokai district of central Japan. Velocity models show that the descending Philippine Sea slab and that low-velocity zones that are associated with volcanism are well delineated.
 1. The Philippine Sea slab comprises an oceanic crust and a high-velocity mantle slab. The oceanic crust shows higher velocities than the surrounding island arc crust at depths shallower than 25 km and shows lower velocities than the surrounding mantle at depths of 25–70 km. At depths greater than 70 km, the oceanic crust shows higher velocities again because it transformed to eclogite at this depth.
 2. The thickness of the descending Philippine Sea slab beneath the Kanto district differs from that located beneath the Tokai district. Their respective thicknesses are about 60 and 25 km. There is a stepwise change of the thickness east of the Izu peninsula. Two alternative explanations exist for the thick Philippine Sea slab. One is that the thick Philippine Sea slab was thick before subduction. Another is that the Philippine Sea slab was thickened after subduction. The actual reason remains unknown, but some relationships exist between the thick slab and Tertiary volcanic activities.
 3. Low-velocity anomalies associated with volcanism were detected. They are located above and west of a 110 km depth contour of the Wadati-Benioff zone, rather than on the backarc side of the volcanic front. The low-velocity anomalies between the volcanic front and the 110 km depth contour of the Wadati-Benioff zone at depths of 30–100 km might be remnants of Tertiary volcanic activities because Tertiary volcanic rocks have been found on the surface of this area.
 We began this study during one author's (S.K.) stay at NIED as a JST postdoctoral researcher. Our special thanks are extended to the staff members of NIED for providing us with data. We are very grateful to K. Koketsu for allowing us to use the seismic ray tracing codes. We also thank T. Seno, S. Maruyama, Y. Okada, K. Kasahara, M. Ukawa, S. Kodaira, K. Fujioka, and K. Aoike for useful discussions and Y. Fukao, Y. Kaneda, and D. Suetsugu for encouraging us and critically reading this manuscript. The comments from two anonymous reviewers and an associated editor were quite helpful to improve the manuscript. All figures were prepared using Generic Mapping Tools (GMT) [Wessel and Smith, 1991].