I'd be cautious as dismissing something as a bad idea because it is slow. If it is a part of the code that does not take much time to execute then the slowness is irrelevant. I just used the following code:
for (ic in 1:(dim(centroid)[2]))
{
cluster[[ic]]=matrix(,nrow=2,ncol=0)
}
# code to identify cluster=pindex[ip] to which to add the point
if(pdist[ip]>-1)
{
cluster[[pindex[ip]]]=cbind(cluster[[pindex[ip]]],points[,ip])
}
for a problem that ran in less than 1 second.