I love sampling algorithms. Here's the sampling algorithm that I find most magical. We want to generate a subset of {1, 2, ..., n} of size k.

def floyd(n, k): s = set() for i in range(n - k + 1, n + 1): t = random.randint(1, i) if t in s: s.add(i) else: s.add(t) return s

I learned about this algorithm the canonical way all good algorithm lore is imparted; one of my old coworkers wrote this comment in some code I was looking at once:

// Use Robert Floyd's algorithm to generate numSplits distinct integers // between 0 and len(splitKeys), just because it's so cool! chosen := make(map[int]struct{}) for j := len(splitKeys) - numSplits; j < len(splitKeys); j++ { t := fit.rng.Intn(j + 1) if _, alreadyChosen := chosen[t]; !alreadyChosen { // Insert T. chosen[t] = struct{}{} } else { // Insert J. chosen[j] = struct{}{} } }

Unlike, say, Reservoir Sampling or something, this algorithm sort of resists an easy intuition to me. The branch makes it weird and makes a normal combinatorial, or visual argument tricky. In fact, I find even a casual intuition hard to grasp. Here are two possible ways to think about what makes it work.

First Perspective

Here's one way to think about it: each iteration of the algorithm goes from a uniformly distributed k-subset of [1..n] and an element of [1..n+1] to a uniformly distributed k+1-subset of [1..n+1]. If we can show that the map from those k-subsets and that sample is k+1-to-1, then we've shown that the resulting set is also uniformly distributed.

So, we need to show that each possible k+1-subset S of [1..n+1] has exactly k+1 inputs that map to it:

• sets containing n+1 come from the k-subset without n+1, and any of the elements of S, and
• sets not containing n+1 come from starting with one of its k-subsets and adding the missing element.

And that's it; the result is uniformly distributed.

This is not so bad, I think it basically makes sense to me. But I also learned of another way of thinking about this algorithm while trying to write this post.

Second Perspective

This one comes from Paul, and it notes a connection between this algorithm and the Fisher-Yates shuffle for permuting an array uniformly at random. Specifically, the "upward" variation that looks like this:

def shuffle(a): for i in range(0, len(a)): j = random.randint(0, i) a[i], a[j] = a[j], a[i]

This builds up a permutation incrementally. After each step, we have a permutation on [0..i], then we augment that and turn it into a permutation on [0..i+1].

Imagine plucking off the last k elements of an array we just shuffled, then that's a uniformly random sample of k elements. The observation here is that Floyd's algorithm is just doing the last k swaps here, directly.

If you call the last k elements of the array the "tail," then for each element i in the tail originally, either we:

• swap i with j that's in the tail, in which case i will stay in the tail forever and be part of the sample, or
• swap it with j not in the tail, then j will live in the tail forever and be part of the sample.

These two cases correspond directly to the two branches of Floyd's algorithm and we wind up with the same sample following them both (up to some reordering).