For diffusion-weighted magnetic resonance imaging and under circumstances where patient movement can be modeled as rigid body motion, it is shown both theoretically and experimentally that translations and rotations produce phase errors which are zero- and first-order, respectively, in position. While a navigator echo can be used to correct the imaging data for arbitrary translations, only when the diffusion gradient is applied in the phase encode direction is there sufficient information to correct for rotations around all axes, and therefore for general rigid body motion. Experiments in test objects and human brain imaging confirm theoretical predictions and demonstrate that appropriate corrections dramatically improve image quality in vivo.