Rolling sum with varying window sizes

Question

I am seeking the fastest way of computing the rolling sum with changing window size. I use the following code, but for vectors of length 1M, it is far too slow.

Thanks

set.seed(1)
n = 10L
x = runif(n)
window = pmin(sample(1:10, n, TRUE), n:1-1)
s = function(x, w){
  n = length(x)
  out = rep(NA, n)
  for(i in 1:n){
    k = w[i]
    out[i] = sum(x[i:(i+k)])
  }
  out
}
s(x, window)
# [1] 2.11869372 1.85318505 4.87750614 3.61375247 3.39644499 3.19476306 2.29637338 1.35169811
# [9] 0.69090031 0.06178627

You could pre-compute `cs = cumsum(x)` and subtract the respective ranges `cs[window + seq_along(x)] - c(0, cs[1:(length(x) - 1)])` — alexis_laz, Jun 25 '16 at 11:03

G. Grothendieck · Accepted Answer · 2016-06-25T13:09:48.773

3

Try this:

en <- seq_along(x) + window   # end positions
cum <- cumsum(x)
cum[en] - cum + x

edited Jun 25 '16 at 13:09

answered Jun 25 '16 at 12:48

G. Grothendieck

254,981
17
203
341

Great solution. Thanks. – ExperimenteR Jun 25 '16 at 13:48

Rolling sum with varying window sizes

1 Answers1

Linked