Behind the sedimentary Ryukyu arc lies the Okinawa Trough, whose termination is located at the tip of the Ilan plain (northern Taiwan), just above the Ryukyu slab edge. The present-day active volcanic front is located 80–100 km above the Ryukyu slab and extends from Japan to Kueishantao Island, an islet situated 10 km offshore the Ilan plain. Between December 1990 and May 1999, 3370 earthquakes recorded in northern Taiwan by 65 seismic land stations were used to determine the three-dimensional Vp and Vs velocity structures and Vp/Vs ratios. A low Vs but high Vp/Vs sausage-like body, ∼30 km in diameter, lies within the Eurasian mantle wedge, on top of the western Ryukyu slab extremity, at depths ranging between 20 and 100 km. We suggest that the H2O-rich component formed by dehydration processes from subducting sediments, oceanic crust, and serpentinized mantle above the Ryukyu slab and along the vertical portion of the slab edge might explain the presence of the sausage-like body. A low Vs but high Vp/Vs channel rises obliquely from the sausage-like body at a depth of 40 km in direction of the andesitic Kueishantao Island. We propose that the H2O-rich component and/or melt rise up from the sausage-like body and interfere with the Okinawa Trough back arc basin magmas formed in the upper mantle/lower crust. Then magmas propagate upward within the upper brittle crust through veins and/or narrow conduits.
 Most arc volcanic rocks are derived from melting of the mantle wedge induced by hydrous fluids released during dehydration reactions in the subducted oceanic lithosphere [Arculus, 1994; Gill, 1981]. Relatively primitive basalts, rhyolites, and andesites have been found in the middle and southern Okinawa Trough (OT) back arc basin [e.g., Shinjo et al., 1999], where is located the present-day active volcanic front (Figure 1). On the basis of velocity tomographic results under northeastern Japan, Wyss et al.  have demonstrated that volcanoes of the volcanic front are linked to a source of magma and fluids originated at the top of the subducting slab and reaching the surface through an oblique mantle and crustal pathway. This observation suggests a strong relationship between the generation of magmas in the OT and the subducting slab. The purpose of this paper is to define the Vp, Vs velocity structures and the Vp/Vs ratios in the area of the western termination of the Ryukyu slab and to provide tomographic images in order to characterize the geometry of melt and/or fluid flow near the slab edge and the magmatic pathway which results in the formation of the Kueishantao Island.
 Taiwan Island is located at the intersection of two converging systems: the Ryukyu subduction zone which results from the subduction of the Philippine Sea (PH) plate beneath Eurasia (EU) and the Manila subduction zone which results from the subduction of EU beneath PH. In this complex geodynamical context, the Ryukyu subduction zone, which extends from Japan to Taiwan, terminates westward beneath northeastern Taiwan (Figure 1). Since early middle Miocene (15 Ma), the continuous oblique N307° subduction of the PH plate beneath EU with respect to the mean N060° direction of the Eurasian margin involves a westward propagation of the Ryukyu slab with respect to EU at a mean velocity of about 4.5 cm/yr [Sibuet et al., 2002]. On the basis of numerous available high-quality earthquake locations and focal mechanisms in Taiwan [e.g., Kao and Chen, 1991; Kao et al., 1998], the Ryukyu slab geometry has been already determined [e.g., Font et al., 1999], but its western end is poorly defined. Earthquake data are always presented as a series of parallel vertical cross sections more or less parallel to the Ryukyu slab edge and do not clearly show the westward disappearance of slab earthquakes [Font et al., 1999; Kao and Rau, 1999; Tsai, 1986]. Similarly, previous tomographic studies have clearly imaged the Ryukyu slab under Taiwan [Lallemand et al., 2001; Rau and Wu, 1995; Roecker et al., 1987] but not precisely the western termination of the slab. However, the triangular shape of the southwestern OT and the location of its tip (western end of the Ilan Plain) just above the Ryukyu slab termination suggest that the overlying OT back arc basin simultaneously propagated westward at the same velocity than the Ryukyu slab [Sibuet and Hsu, 2004].
 The volcanic front (dashed gray line in Figure 1) extends from Japan to Taiwan. From Kyushu to Okinawa Island, it coincides with small subaerial active volcanoes located about 25 km west of the ancient volcanic arc. Southwest of Okinawa Island, it follows numerous seamounts associated with high-magnitude magnetic anomalies [Hsu et al., 2001], then the cross-back-arc volcanic trail, which consists of a cluster of about 70 seamounts located west of 123°E longitude, and finally ends 10 km offshore Taiwan at Kueishantao Island [Chung et al., 2000; Sibuet et al., 1998], an island located in the axis of the nascent OT. The volcanic front is located 80–100 km above the subducted slab. 80 km is the minimum slab depth required for the emplacement of arc magmatism [Gill, 1981; Tatsumi, 1986], explaining why the volcanic front is located within the OT, north of the Ryukyu nonvolcanic arc where only sedimentary islands outcrop. Rocks from the present-day volcanic front are predominantly andesites, although they range from basalts to rhyolites both in the northern Ryukyu arc [Daishi, 1992; Nakada, 1986] and in the southwestern OT [Chung et al., 2000; Shinjo, 1999] and andesites in Kueishantao Island [Chen et al., 1995; Chung et al., 2000]. In fact, the location of the volcanic front in the axis of the southwestern OT back arc basin favors the emplacement of back arc basin volcanic rocks with a large involvement of a subduction component [Shinjo, 1999].
 In the following sections, we will try to show how tomographic images, even if data are restricted to the western Ryukyu slab edge and Kueishantao Island, can provide useful images concerning the locations of magmas and fluids and their ascending pathways through the overlying EU mantle and crust.
 Between December 1990 and May 1999, 3370 earthquakes occurring north of 23.5°N were extracted from the Central Weather Bureau database (Figure 2). The selection of earthquakes is based on the following criteria: (1) Earthquakes must have occurred within the seismic recording net of stations; (2) their magnitude must be larger than 3.0; (3) P arrivals must be recorded at least by 10 stations; and (4) earthquakes with more than four bad quality records (weight 4 in the sense of the “SIMUL 2000” program [Eberhart-Phillips, 1986, 1990; Thurber, 1983, 1993; Thurber and Eberhart-Phillips, 1999]) are removed. In other words, earthquakes whose events display time residuals larger than 50 s after the first time inversion were removed. The “flexible gridding” method [Thurber and Eberhart-Phillips, 1999] increases the fine-scale resolution and prevents an inhomogeneous ray path distribution. Different gridding space intervals were applied in function of the ray path density. In the volumes of high ray path density, the use of fine gridding interval increases the fine-scale resolution. In the regions of poor checkerboard test resolution, values of the slave grids (gray circles) are identical to values of the adjacent master grids (stars) (Figure 3). The spacings for master and slave nodes are 18 and 6 km respectively. In addition, the linkage of slave and master grids helps to recover a reliable smoothed structure in the volumes of low ray path distribution.
 A total of 83,107 P wave and 44,412 S wave arrival times recorded by 65 seismic land stations were selected in this study. In the velocity model of Chen , the island of Taiwan was divided in three areas (west, northeast, and southeast Taiwan). The average one-dimensional (1-D) model of the northeast area was used as the initial 1-D starting model in our inversion. The “SIMUL 2000” program was applied to inverse the Vp, Vs, and Vp/Vs structures. A “pseudobending” synthetic ray path method was used to calculate travel times [Um and Thurber, 1987]. After parameter decomposition, a damped least squares technique was applied to inverse the P and S wave velocity structures. The ratios of 3-D Vp and Vs models to determine the 3-D Vp/Vs structure can lead to severe artifacts in areas where Vs is badly resolved [Eberhart-Phillips, 1990; Thurber, 1993]. Thus, instead of computing directly the Vp/Vs structure from Vp and Vs values, we have used the Vp and Vs residual time arrivals to calculate the Vp/Vs ratios. Tomographic results are displayed in percentage of variations with respect to the average value calculated at a given depth as proposed by Thurber and Eberhart-Phillips .
 The resolution R is defined by the equation mest = Rmtrue given by Menke  where mest is the estimated set of model parameters, mtrue is the true but unknown set of model parameters, and R is the model resolution matrix. mtrue is the solution of G mtrue = dobs, where G is the data kernel in analogy with the theory of integral equations and dobs are the arrival times. The model resolution matrix R shows how mest is close from the true solution mtrue. The quality of resolution depends on the similarity between the model resolution matrix (R) and the identity function (I). If R = I, each model parameter is uniquely defined. The diagonal elements of the resolution matrix represent the resolvability of the model (resolution). When the diagonal elements are close to 1.0, the estimated parameters (mest) mostly correspond to the real data. According to the trade-off curve established from the data variance and the model variance, the damping values were chosen as 45 for Vp and as 40 for Vp/Vs inversion in order to minimize the data and the model variances.
2.2. Tomographic Inversion and Results
Figure 3 shows the results of P waves checkerboard resolution tests. We assigned positive and negative velocity perturbations of ±3% alternately to the grid nodes and calculated travel times for this model to produce synthetic data. The synthetic data were then inverted with an initial model of zero velocity perturbation. The results of P waves checkerboard resolution tests are excellent in the eastern and northeastern portions of the data volume for shallow depths (15 to 60 km), and also for larger depths but only in the restricted northeastern corner of the data volume. Figures 4 to 6 present P and S wave and Vp/Vs ratio perturbation tomographic images for 12 depth slices ranging from 2 to 125 km. Figure 7 shows, as an example, the Vp data resolution, which is very similar to the Vp/Vs data resolution (not shown here). Vp/Vs values higher than 1.78 have been contoured in Figure 6 and transferred on the other figures. Except for shallow areas, the resolution is of good quality. In particular, the area of the Ryukyu slab edge is properly imaged from 7 to 125 km (Figure 7). The hypocenter relocation results show an average shift of 2.6 km in the horizontal plane and 0.4 km in depth, but this information is not used in this paper.
 On Vp/Vs tomographic slices (Figure 6), a series of high-value anomalies located onshore, along the eastern coast of Taiwan progressively deepen to the north from 20 to 125 km deep, as highlighted by contoured Vp/Vs values higher than 1.78. These high Vp/Vs anomalies are located in the area of the western termination of the Ryukyu slab, just above the subducting slab. On S wave tomographic images (Figure 5), the Vp/Vs high-value anomalies are systematically associated with low Vs perturbation anomalies. However, on P wave tomographic images (Figure 4), such a correlation with low Vp perturbation anomalies is not obvious.
 Two Vp, Vs and Vp/Vs vertical profiles (located on Figure 8) have been extracted (Figure 9). Profile 1 is located on the EU continental crust, west of the Ryukyu slab termination, whereas profile 2, located 20 km east of profile 1, has been chosen along the trend of high Vp/Vs anomalies. On profile 1, located about 10 km west of the Ryukyu slab edge, both the slab and the high Vp/Vs velocity body have disappeared. Tomographic data obtained with less data (1197 earthquakes) and a similar analytical method by Rau and Wu  show similar trends along profile 2 (Figure 9). On the two data sets, the continental crust about 35 km thick and the Ryukyu slab are well imaged. However, the slab is better imaged in our data set because the tomographic gradient between 60 and 120 km corresponds to the earthquake epicenter envelope, which could be considered as an index of quality of our tomographic results. The high Vp/Vs body is 30 km thick and is located above the Ryukyu slab (profile 2, Figure 9), within the EU continental crust and mantle. It continuously extends from a depth of 15 to 80 km for its top and 30 to 120 km for its floor. It was already but vaguely identified on Rau and Wu's  profile shown in Figure 9. The floor of this body corresponds exactly to the top of the Ryukyu slab, underlined by the upper envelope of earthquake locations. The horizontal width of the body is ∼30 km and remains constant throughout depth (Figure 8). Consequently, Vp/Vs data show the presence of a continuous body ∼30 km wide and 30 km thick extending from mean depths of 20 to 100 km and located just above the edge of the Ryukyu slab. On profile 2 cross sections of Vs perturbation anomalies, at depths deeper than 15 km, this body corresponds exactly to a low-velocity structure, which extends down to the same depth than for the Vp/Vs body (Figure 9).
Figure 10 shows the 3-D geometry of the sausage-like body as well as a continuous and relatively high Vp/Vs perturbation anomaly, which rises northerly in direction of the Ilan plain axis from a depth of 40 km to the subsurface. Then, the feature turns easterly at subsurface depths, along the back arc basin axis, in direction of Kueishantao Island. Low Vs perturbation anomalies are also associated to this trend, except between depths ranging from about 7 to 13 km where higher Vp and Vs perturbation values are observed (Figure 9). The presence of an oblique feature reaching Kueishantao Island through the mantle wedge confirms that the magmatic origin of Kueishantao andesites is linked to the slab itself and that fluids and/or melt rise along an inclined path as demonstrated by Wyss et al.  for arc volcanoes in northern Japan.
 To summarize, we have identified two continuous bodies characterized by high Vp/Vs and low Vs: (1) one with a sausage-like shape, ∼30 km in diameter and located within the EU mantle wedge, on top of the western Ryukyu slab termination, between depths of 20 and 100 km, and (2) the other one rising northeastward and upward from the sausage-like body from a depth of 40 km in direction of Kueishantao Island. However, high Vp and Vs interrupting bodies are observed between depths of about 7 to 13 km as also noticed beneath the Japanese arc volcanoes [Nakajima et al., 2001b].
2.3. Interpretation and Tectonic Implications
 Magmatism associated with subducting plate edges or slab tears has been reported several times. Abrupt cutoffs of the deep seismicity offshore Sicily [Frepoli et al., 1996] give the location of the southern edge of the Ionian subducting lithosphere. Above the slab edge, a few volcanic edifices are spanning from the Etna (Sicily), located a few tens km above the slab edge, to Ustica Island located 400 km above the slab edge. They are characterized by an OIB-type activity which displays a peculiar geochemistry compatible with a slight mantle source contamination from subduction-related fluids [Beccaluva et al., 1982; Marani and Tra, 2002], allowing Marani and Tra  to suggest the existence of upwelling asthenospheric flow above the slab edge.
 High-resolution 3-D tomographic images obtained in northern Honshu (Japan) show that low Vp, low Vs and high Vp/Vs are distributed in the upper mantle and crust along the volcanic front, with large crustal anomalies beneath individual volcanoes [Nakajima et al., 2001a, 2001b]. These anomalies rise from the top of the subducting slab (here at a depth of 150 km) along inclined mantle paths to the volcanoes, except between 8 and 18 km where a zone of high Vp is observed. Wyss et al.  suggest that the inclined paths are associated with positive thermal anomalies and follow the expected mantle flow trajectories in the wedge located above the slab [Tamura et al., 2002].
Watanabe  shows that the melt fluid fraction significantly decreases Vp and Vs velocities; on the contrary, the effect of seawater contributes to the Vp decrease but does not have any effect on Vs. Nakajima et al. [2001b] and Takei  also calculated Vp, Vs, and Vp/Vs variations when cracks in matrix are filled with H2O or melt. They conclude that low Vp, low Vs, and high Vp/Vs in the lower crust and upper mantle are caused by partial melting and an H2O-rich component because of a need of a too small aspect ratio of cracks in this later case.
 What physical mechanism could explain the existence of the continuous sausage-like body from 20 to 100 km depth above the Ryukyu slab edge? The continuous 8 cm/yr subduction velocity of the Ryukyu slab in the N307° direction with respect to the EU lithosphere could generate some localized heating by friction of the two plates along the western vertical edge of the slab. However, Rüpke et al.  suggest that shear stresses and shear heating, though largely unconstrained, are weak at the boundary between the slab and the adjacent mantle and lithosphere. Thus friction heating might be negligible. If the 800°C minimum required temperature to generate some melt occurs at depths larger than 80 km, a slight increase in temperature would not significantly shallow the production of melt. It is thus impossible to explain by friction heating the extension of the sausage-like body up to 20 km. Therefore a general mechanism, as an upward migration of fluids generated by deep dehydration, might exist to account for the fluid enrichment. Fluid release occurs above the slabs at depths <20 km from subducting sediments, at intermediate depths (20–100 km) from sediments and oceanic crust and at depths >100 km from oceanic crust and serpentinized mantle [Rüpke et al., 2004]. Hacker et al.  argue that intermediate depth earthquakes only occur where hydrous minerals are predicted to be present, implying a causal link between dehydration reactions and seismicity. The slab seismicity being more developed beneath the sausage-like body than along the rest of the slab [Kao et al., 2000], it suggests that an excess of the H2O-rich component might be expected at the slab border and released along the vertical edge of the Ryukyu slab. Another possibility would be the upwelling of hot underlying PH lithospheric mantle around the Ryukyu slab edge due to mantle flow initiated within the ductile PH mantle by the westward PH motion with respect to EU.
 To summarize, we suggest that the H2O-rich component and/or melt would be conveyed around the edge of the slab to form the sausage-like body located above the Ryukyu slab edge. Then the upward migrating H2O-rich component derived from the sausage-like body might interact with the overlying EU lithosphere and enhance the generation of magmas located in the EU upper mantle/lower crust [Stolper and Newman, 1994], in the axis of the OT back arc basin.
 High-Mg (5%) andesites have been recovered in Kueishantao Island [Chen et al., 1995; Chung et al., 2000]. Following Kelemen  and Shinjo , they might result from the reaction between EU mantle peridotite and the ascending liquid originated from the sausage-like body. For the first time, we have imaged that the feeding of a volcano seems to be directly connected to the H2O-rich component and/or melt produced at the slab edge. The location of Kueishantao Island in the OT axis, even if back arc basin extension started there very recently (<1 Ma) [Sibuet and Hsu, 2004] and is considerably reduced close to the OT termination, would favor the ascent of basaltic magma formed in the lower crust and/or upper mantle and already contaminated by the deep H2O-rich component. Magmas would rise up along upper crustal normal faults developed within the back arc basin. Kueishantao andesites have very high 87Sr/86Sr and low 143Nd/144Nd, which lead Chen et al.  to propose that the Kueishantao magmas were substantially contaminated by the upper crust. As the most recent Kueishantao andesites erupted recently (8000 years) and hydrothermal activity is observed today on Kueishantao Island [Lee et al., 1998], the feeding channel is probably still active along the continuous low Vs and high Vp/Vs channel. Yet, the high Vp and high Vs perturbation anomalies from about 7 to 13 km might suggest that the H2O-rich component and/or melt are restricted to depths larger than 13 km and only reach lower crustal levels. We also suggest that magmas propagate upward and eastward within the upper brittle crust through magmatic veins and/or narrow magma conduits because the crust of the back arc basin thins and is more tectonized eastward.
 Such a model is illustrated in the sketch of Figure 11. We suggest that the same mechanism also applies for the Ryukyu slab tear located along the 123.3°E meridian (Figure 1). A large cluster of volcanoes, called the cross-back-arc volcanic trail (CBVT) [Sibuet et al., 1998] consists of basalts, rhyolites, and andesites. It is located in the axial part of the southern OT, at 123°E longitude and seems to be linked obliquely with the slab tear. We suggest that an excess of H2O-rich component and/or melt might be formed at the slab tear and might increase the melt flux [Lin et al., 2004], involving the same process than for the sausage-like body. Both are conveyed obliquely in direction of the uppermost mantle and lower crust CBVT magmas. From there, after geochemical interactions with the H2O-rich component coming from below, basaltic magmas rise up along normal back arc basin faults, in the CBVT area.
 The main conclusions of this study are as follows:
 1. 3370 earthquakes recorded in northern Taiwan by 65 seismic land stations between December 1990 and May 1999 were extracted from the database of the Central Weather Bureau in Taiwan. The 3-D Vp and Vs velocity structures and Vp/Vs ratios were determined by inversion of P wave arrivals and P-S arrival residuals by using a “pseudobending” ray path method. A low Vs but high Vp/Vs sausage-like body was identified above the western extremity of the Ryukyu slab, from depths ranging from 20 to 100 km. This sausage-like body, identified for the first time in this type of environment, is 30 km wide and 30 km thick. We suggest that the H2O-rich component formed by dehydration processes from subducting sediments, oceanic crust and serpentinized mantle above the Ryukyu slab and along the vertical portion of the slab edge might explain the presence of the sausage-like body.
 2. A low Vs but high Vp/Vs channel rises obliquely from the sausage-like body at a depth of 40 km in direction of the andesitic Kueishantao Island. The presence of high Vp and high Vs perturbation values from about 7 to 13 km might suggest that the H2O-rich component and melt are restricted to depths larger than 13 km and only reach lower crustal levels. We propose that the H2O-rich component and melt rise up from the sausage-like body and interfere with the OT back arc basin magmas formed in the upper mantle/lower crust. Then, magmas propagate upward within the upper brittle crust through veins and/or narrow conduits not imaged in our tomographic results. The high-Mg andesites are likely to be the product of reaction between EU mantle peridotite and the ascending liquid originated from the sausage-like body.
 3. A similar mechanism might explain the presence of an excess amount of basaltic and rhyolitic volcanoes located above the Ryukyu slab tear located at 123.3°E longitude.
 The GMT software package was used to draw some of the figures [Wessel and Smith, 1991]. This work is part of an ongoing cooperative project between France and Taiwan encouraged and supported by the French Institute of Taipei and by the National Science Council, Taiwan. Constructive discussions with S.-L. Chung, L. Ciron, E. Debayle, L. Dosso, S. Lallemand, C.-S. Liu, R. Maury, W. Roest, and R. Shinjo are acknowledged. We also acknowledge the very constructive reviews of G. Abers, A. Hasegawa, H. Kao, and C.-S. Lee.