8
$\begingroup$

I want to calculate/evaluate the convolution

$$g(x)=\int_D f(x-t) \phi(t) dt,$$

where $f$ is a density and $\phi$ is a smooth function with compact support $D$. The convolution is not available in closed-form and I need to integrate it numerically. My question is: Is there an efficient way to do this? I want to implement it in R, so, I would like to see if there is a better way than using the command integrate().

Improve this question
$\endgroup$
5
  • 2
    $\begingroup$ Depending on circumstances I usually either discretize to a large power of 2 bins, and use the fast fourier transform (?fft) or use convolve. The fft approach takes a bit more work to set up but is better if you need to convolve with something several times. Sometimes it takes a while to figure out the right argument settings with convolve. $\endgroup$
    – Glen_b
    Commented Nov 5, 2014 at 13:52
  • $\begingroup$ @Glen_b Thanks. For a univariate function $f$, I think the direct integration may be faster, then. $\endgroup$
    – Cook
    Commented Nov 5, 2014 at 13:53
  • 2
    $\begingroup$ You asked for an efficient way - fft is really fast; it just requires a little bit of setup (binning, padding with zeros). $\endgroup$
    – Glen_b
    Commented Nov 5, 2014 at 13:55
  • $\begingroup$ @Glen_b Yeah, I agree the fft is really fast, but the previous step may slow down the process. I will compare both methods, anyway. Thank you. $\endgroup$
    – Cook
    Commented Nov 5, 2014 at 13:56
  • $\begingroup$ I recall using convolve for this purpose several times. Simple working examples appear at stats.stackexchange.com/a/41263, stats.stackexchange.com/a/41255, and stats.stackexchange.com/a/49444. $\endgroup$
    – whuber ♦
    Commented Nov 5, 2014 at 15:47

1 Answer 1

Reset to default
11
$\begingroup$

Have you taken a look at dedicated R packages for that? Like convolve, https://stat.ethz.ch/R-manual/R-devel/library/stats/html/convolve.html

Improve this answer
$\endgroup$
2
  • $\begingroup$ Thanks. If I properly understand, this package is for convolving numeric sequences, rather than two functions. Am I missing something? $\endgroup$
    – Cook
    Commented Nov 5, 2014 at 13:51
  • 2
    $\begingroup$ Well, I've not used that package, but if you're trying to do the convolution in R, then you'll basically have two numeric sequences which will be the values from your functions, wouldn't they? I don't think R lets you work with the function definition. $\endgroup$
    – jmnavarro
    Commented Nov 5, 2014 at 13:53

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.