Predictability of particle motion in the ocean over a time scale of one week is 
studied using 3 clusters of buoys consisting of 5-10 drifters deployed in the 
tropical Pacific Ocean. The analysis is conducted by using three techniques with 
increasing complexity: the center of mass of the cluster, advection by 
climatological currents, and a new technique, which relies on the assimilation 
of both velocity and position data from the surrounding drifters into a Markov 
model for particle motion.

The results indicate that cluster predictability can be characterized using the 
data density Nd, defined as the number of drifters over an area scaled by the 
mean diameter of the cluster. The data density Nd decreases along the drifter 
trajectories due to the tendency of particles to disperse by turbulent fluid 
motion. In the first regime, which corresponds to the period after the release 
of drifters in a tight cluster when Nd>>1 drifter/degree2, the center 
of mass and the data assimilation methods perform nearly equally well, and both 
methods yield very accurate predictions of drifter positions with rms prediction 
errors less than 15 km up to 7 days. When a cluster starts to disperse, i.e., in 
the regime where Nd>1 drifter/degree2, the data assimilation 
technique is the only method that gives accurate results. Finally, when Nd<<1 
drifter/degree2, no method investigated in this study is effective. 
Uncertainties in the knowledge of initial release positions and the frequency of 
data assimilation are found to have a strong impact on the prediction accuracy.