Monday, September 19, 2011

Follow-up: Sampling a discrete distribution

This is a follow-up to my last post, on a puzzle about "Sampling a discrete distribution" (see my comment there for the solution I originally thought of).


As an anonymous commenter (Rex?) points out, a nice elementary solution that has been known for a long time. I admit to not knowing this solution ahead of time, and use the fact that it's not really my field as an excuse. Here I want to summarize this solution for educational purposes.

Imagine each sample in the distribution is a vase containing liquid proportional to the probability of the outcome.

We classify these vases into 3 types:
LO —  probability less than 1/n
GOOD –  probability precisely 1/n
HI –  probability greater then 1/n

In the ideal case when every vase is GOOD, we are done (uniform sampling). If not, we  move towards this ideal case by inductively applying the following subroutine:

** Pick a LO vase (say, with x amount of liquid) and a HI vase (with y amount of liquid).

Grab a new vase and fill it up to the 1/n mark: pour all the liquid from the LO vase, and (1/n)-x liquid from the HI vase (remember that a HI vase has liquid y>1/n, so this is certainly possible). Note that the HI vase may now become LO, GOOD, or stay HI.

/end of **

Observe that the process terminates in n steps, since at each step we increment the number of GOOD vases.  If we maintain 3 linked lists with samples of each type, each step can be done in  O(1) time, so O(n) time total.

At the end all vases are GOOD so we use uniform sampling. If we get a random x in [0,1], we use ⌊x·n⌋ to indicate that vase and x·n mod 1 to pick a molecule of  water inside that vase. If we trace back the original vase from where this molecule came, we have sampled according to the initial distribution.

Now we describe how to "undo" the water pouring at query time, i.e. tracing the origin of each water molecule. Observe that operation (**) is never applied to a GOOD vase. So when a vase becomes GOOD, it has reached final form. But the water in a GOOD vase can only come from two distinct vases. So we can essentially draw a water-mark on the vase, to separate the LO water from HI water and store pointers to the original vases whose water got poured into the GOOD vase. Then the sampling algorithm only need one comparison, i.e. it compares x·n mod 1 to the mark on the vase ⌊x·n⌋, and follows a pointer from this vase to the input vases from which water was poured. Note that only one pointer is followed, since a GOOD vase is never touched again by operation (**). Constant time!

Friday, September 16, 2011

Sampling a discrete distribution

The following cute question came up at lunch today (all credit goes to Aaron Archer).


Informal challenge: You are given a discrete distribution with n possible outputs (and you know the probability pi of each output), and you want to sample from the distribution. You can use a primitive that returns a random real number in [0,1].

This was the way the problem came to me, and the goal was to find a very fast practical solution. Think of n as around 106 and use a regular machine as your machine model (real number means a "double" or something like that).


Formal challenge: To formalize this completely for a bounded-precision computer, we can say that we work on a w-bit machine and probabilities have w-bit precision. The distribution is given by n integers xi, and these values sum to exactly 2w. Output "i" should be produced with probability exactly xi / 2w. You are given a primitive that produces a uniformly distributed w-bit number.

The challenge is to get O(1) worst-case time. (NB: some fancy data structures required.)