There are several strategies for building k-d trees. The offline method (1) accumulates all the data in an array, (2) finds the best dimension to discriminate on, namely, the one with the widest range, (3) finds the best value of that dimension to discriminate on, namely, the median value in that dimension (typically using the QuickSelect algorithm with Lomuto's partitioning method), (4) separates the data into two subarrays based on that discriminant, (5) recurses back to step 2 on each subarray. Recursion terminates when an array has size b or smaller. One can also devise online methods that add to existing trees.