If it is useful, then a practical poin is how to generate good discrepancy. Chazelle describes a set of concepts which are not really algorithms.

My thought about an effective algorithm would be something like this:

We keep a running set of queues in the form of a geometric hash for the bins we are trying to fill and use a pseudo-RNG with rejection to add points to the working bins, internally structured as priority queues. When the grid changes, we can mix old & new candidates according to the new priority structure. Priority could be based on rejection (or acceptance) for spectral clusters that we care about in each cubicle.

Thoughts? Detail?