Style and variation are two vital components of human motion: style differentiates between examples of the same behavior (slow walk vs. fast walk) while variation differentiates between examples of the same style (vigorous vs. lackadaisical arm swing). This paper presents a novel method to simultaneously model style and variation of motion data captured from different subjects performing the same behavior. An articulated skeleton is separated into several joint groups, and latent variation parameters are introduced to parameterize the variation of each partial motion. The relationships between user-defined style parameters and latent variation parameters are represented by a Bayesian network that is automatically learned from example motions. The geostatistical model, named universal Kriging, is extended to be a style-and-variation interpolation to generate partial motions for all joint groups. Experiments with sideways stepping, walking and running behaviors have demonstrated that the motion sequences synthesized by our method are smooth and natural, while their variations can be easily noticed even when their input style parameters are the same.