A potential solution for this is just sample the burst, i.e. only a random set of the non-predicted column can burst. If the size of that set is bigger than the sparsity, you still will connect the sequence. In my implementation, that approach seems to be working just fine.