A vertical tomogram is constructed for the mantle between Tonga and Hawaii. Using a complete Gaussian-Bayesian approach, we inverted a data set comprising 304 ScS reverberation travel times with Fr¿chet kernels computed by the paraxial ray approximation and 1122 frequency-dependent phase delays of turning and surface waves with Fr¿chet kernels calculated by a normal-mode coupling algorithm. The model parameters include shear speed variations, perturbations to shear wave radial anisotropy in the uppermost mantle, and the topographies of the 410- and 660-km discontinuities. Tests demonstrate that the data set resolves lateral structures in the upper mantle with scale lengths of several hundred kilometers and greater. The most significant feature in our model is a well resolved regular pattern of high and low shear velocities in the upper mantle. These variations have a horizontal wavelength of 1500 km, a vertical dimension of 700 km, and an amplitude of about 3%, and they show a strong positive correlation with seafloor topography and geoid height variations. The geoid highs correspond to a series of northwest trending swells associated with the major hotspots of the Society, Marquesas, and Hawaiian Islands. Where these swells cross the corridor, they are underlain by high shear velocities throughout the uppermost mantle, so it is unlikely that their topographies are supported by thermal buoyancy. Although chemical buoyancy generated by basaltic volcanism and the formation of its low-density peridotitic residuum can explain the positive correlation between topography and velocity variation at the shallow depth, an additional mechanism is needed to account for the shear velocity pattern at depths greater than 200 km. It is therefore hypothesized that the central Pacific is underlain by a system of Richter-type convective rolls that are oriented subparallel to the absolute plate velocity, have unit aspect ratio transverse to this orientation, and are confined above the 660-km discontinuity. This convection pattern appears to be strongly correlated with the locations of the Tahitian, Marquesan, and Hawaiian hotspots, which raises interesting questions for Morgan's hypothesis that these hotspots are the surface manifestations of deep mantle plumes. ¿ 1998 American Geophysical Union