We analyze strong-motion recordings of the M_s 6.6 Superstition Hills earthquake to determine the timing, location, spatial extent, and rupture velocity of the subevents that produced the bulk of the high-frequency (0.5 to 4 Hz) seismic energy radiated by this shock. The earthquake can be characterized by three principal subevents, the largest ones occurring about 3 and 10 sec after initiation of rupture. Timing relationships between pulses on the seismograms indicate that the three subevents are located within 8 km of each other along the northern portion of the Superstition Hills fault. The two largest subevents display different directivity effects. We apply a tomographic source inversion to the integrated accelerograms to determine the slip acceleration on the fault as a function of time and distance, based on a one-dimensional fault model. The azimuthal distribution of amplitudes for the second subevent can be largely explained by a rupture that propagated about 2 km to the southeast along the Superstition Hills fault at a velocity about equal to the P-wave velocity. An alternative model with rupture propagating to the northeast along a conjugate fault plane can also account for the observed directivity of this subevent, but it is not supported by the aftershock distribution. The third subevent ruptured to the southeast along an 8-km long portion of the Superstition Hills fault at about the shear-wave velocity. This rupture propagation caused the relatively large accelerations and velocities observed in strong-motion records for stations southeast of the hypocenter. The long time intervals between the subevents and their relative proximity to each other indicate a very slow component to the rupture development. The southern half of the Superstition Hills fault did not generate significant high-frequency strong ground motion, although it showed substantial co-seismic surface displacement. The subevents are situated along the same northern portion of the fault where most of the aftershocks are located. The locations of the subevents appear to be controlled by bends in the fault mapped at the surface and by changes in basement structure at depth.