I'm using R 3.6.2 (platform = x86_64-w64-mingw32)
In the convolution of the following two polynomial coefficient vectors, I would expect the first entry to be exactly 1.0, but the convolve
function differs:
g <- c(1, -49, 1155, -17441, 189700, -1583071, 10545901, -57608692,
263063351, -1018546561, 3380085631, -9693547553, 24176423345,
-52691112850)
u <- c(1, -6, 11, -6)
convolve(g, rev(u), type = 'o')
# output
[1] 1.000172e+00 -5.500020e+01 1.460000e+03 -2.491600e+04
[5] 3.073450e+05 -2.920052e+06 2.223567e+07 -1.394361e+08
[9] 7.342188e+08 -3.293898e+09 1.273071e+10 -4.275645e+10
[13] 1.256299e+11 -3.246592e+11 6.402486e+11 -7.246608e+11
[17] 3.161467e+11
Note that the first entry in the result is 1.000172, not 1.0.
Performing the same convolution in Python 3.7.4 delivers the expected answer:
import numpy as np
g = [1, -49, 1155, -17441, 189700, -1583071, 10545901, -57608692, 263063351, -1018546561, 3380085631, -9693547553, 24176423345, -52691112850]
u = [1, -6, 11, -6]
np.convolve(g,u)
array([ 1, -55, 1460, -24916,
307345, -2920052, 22235673, -139436079,
734218840, -3293897685, 12730714010, -42756453616,
125629929970, -324659189789, 640248619213, -724660781420,
316146677100], dtype=int64)
Also when I use the convolveCpp
example from the Rcpp vignette, I get the same results as in the Python above.
Is there a rounding or precision issue with convolve
or the underlying fft
?