To investigate the seismic structure of the Tonga subduction zone, mobile seismic station were temporarily operated on the four main islands of the Tonga group. Data from intermediate-depth ad deep earthquakes were analyzed using P wave arrival times only. To improve on single-event locations as published in bulletins, a joint hypocentral determination method was applied to three groups of deep earthquakes and two groups of intermediate-depth earthquakes. The resulting P wave travel time residuals observed in the Tonga island arc are generally negative and reveal a characteristic pattern. A notable feature is the observation that residuals from deep earthquakes located near 18¿ S are consistently less negative than residuals from deep earthquakes located south of 20.5¿ S. It cannot completely be ruled out that part of this difference is caused by systematic depth errors that may arise as a result of slab propagation effects on the upgoing p leg of pP, but only a minor part of the observed difference can be explained in this way. The major part is probably caused by two factors. First, P waves traveling from the southern part of the Tonga subduction zone to the northernmost stations located in the islands of Vava'u and Ha'apai spend more time in the high-velocity slab than P waves originating in the northern part of the Tonga seismic zone. Second, ray tracing through a number of models of the subduction zone suggests that P waves from deep earthquakes located wouth of 20.5¿ S have their propagation paths close to the high-velocity core of the slab, whereas P waves from deep and perhaps also from intermediate-depth earthquakes located near 18¿ S are likely to miss parts of the high-velocity slab on their way up to the island arc.

This is most pronounced for the seismic station TGA, located in the island of Tongatapu, which is some 30 km further away from the trench than the other stations of the network. The magnitude of residual differences (2--3 s) at TGA observed from different groups of deep earthquakes is best explained by a sharp bend in the dip of the subducting plate in the northern part of Tonga near a depth of 200 km. Ray tracing through such a model shows that P waves from deep earthquakes near 18¿ S traveling to TGA escape the slab at a depth of about 250 km and, from there on, propagate through the mantle overlying the slab. If the interpretation is correct, P wave velocities are indicated that are 5--7% higher in the slab above 250 km than in the overlying mantle. Independent of this interpretation, the magnitude of negative residuals from deep earthquakes south of 20.5¿ S is best matched by velocities in the slab that are uniformly higher by 6% relative to the Herrin model. In another model where P wave velocities vary linearly over horizontal slab width from maximum values in the seismic zone to the velocities in the Herrin model at the slab boundaries, the data are equally well fitted if the velocity anomaly reaches a maximum value of 8% relative to the Herrin model. Allowance for a crustal correction term of -0.5 s reduces these estimates by 1%. Modification of the velocity models to include the case of a velocity anomaly in the slab that decreases linearly with depth to 0% over the range from 400 to 700 km increases the estimates by 1%. The data do not allow a proper assessment of the question of whether seismic velocities approach normal mantle values at depths below 400 km or remain substantially higher than in standard models of the earth. ¿ American Geophysical Union 1987