Fork me on GitHub
We are hiring a PhD student to do research on functional high performance computing, including work on Futhark itself. See here for more details. Source file: variance.fut


A naive way of computing variance of an array where each element is considered equally likely is as follows, which is close to the textbook formula:

def mean [n] (vs: [n]f64) =
  f64.sum vs / f64.i64 n

def variance [n] (vs: [n]f64) =
  let m = mean vs
  let xs = map (\x -> (x-m)*(x-m)) vs
  in f64.sum xs / (f64.i64 n)

Unfortunately, sums of squares can lead to numerical instability. A parallel adaptation of Welford’s Online Algorithm can be used to obtain a more stable variance. The basic idea is a divide-and-conquer parallel algorithm where a subsequence of the data is described by a triple containing:

  1. The length of the subsequence.
  2. The mean of the subsequence.
  3. The estimated variance of the subsequence.

If we have an associative operator for combining two such triples, then we can write it as a map-reduce composition in Futhark. Fortunately, such an operator exists:

def red_op (n_a, m_a, m2_a) (n_b, m_b, m2_b) =
  let n_ab = n_a + n_b
  in if n_ab == 0 then (0, 0, 0)
     else let f_a = f64.from_fraction n_a n_ab
          let f_b = f64.from_fraction n_b n_ab
          let f_ab = f64.from_fraction (n_a * n_b) n_ab
          let delta = m_b - m_a
          let m_ab = f_a * m_a + f_b * m_b
          let m2_ab = m2_a + m2_b + delta * delta * f_ab
          in (n_ab, m_ab, m2_ab)

The only real subtlety is that we need to special-case a length of zero to avoid division by zero. We can now define a numerically stable function for computing variance.

def variance_stable [n] (X:[n]f64) =
  let (_, _, m2) =
    reduce_comm red_op (0, 0, 0) (map (\a -> (1, a, 0)) X)
  in (m2/f64.i64 n)

Note also that this function is a single-pass algorithm.

Thanks to Gusten Isfeldt for this example.