That's similar to the QuickSort algorithm. Recursively break the set into pairs, with one pair having everything < a given value, the other having those >=.
One difference is that QS assumes that the first item in a given (sub)set is a good representative of the median value. Thus, QS is O(N log(N)) on average (typical with random data), but if you're data is already sorted it falls to O(N^2) like a Bubble Sort.
Anyway, you're still stuck with O(N log(N)) complexity on something that's theoretically possible in O(N) for any input dataset.
But I'm still learning functional programming myself, and don't know how to do any better than your proposal.