1
$\begingroup$

How does one do MC for integrand which has dirac-delta function like following:

$I=\int e^{-S(x)} \delta(f(x))dx$

where $x$ would be multi-dimensional, and hence this is a multi-dimensional integral over coordinates such as $x_1,x_2,x_3$ and so on.

I want to do importance sampling. I do have my sampling distribution $g(x)$ which is nice, and closely resembles $e^{-S(x)}$. My major problem is dealing with dirac-delta. I want to do it completely using MC. In other words, I don't want to integrate out dirac-delta analytically, and perform the rest of the integrals over $x_2,x_3$, etc. by MC. Rather, everything is supposed to be done numerically using MC. Does anyone have any idea how to deal with dirac-delta distribution in monte carlo sampling methods?

$\endgroup$
  • 2
    $\begingroup$ Why do you want to do it using Monte-Carlo? $\endgroup$ – user14717 Jun 26 '17 at 15:26
  • $\begingroup$ because it is multi-dimensional, and also the integrand has some poles. how does one avoid poles of integrand? I am thinking of rejecting the samples that create pole when doing MC sampling. $\endgroup$ – t.jun Jun 27 '17 at 8:11
  • $\begingroup$ I don't think the integrand having poles is a good reason to do Monte Carlo, in fact it's a bad reason. The sampling will never hit the pole, nor will it be able to delineate its structure by coming very close to it in a sophisticated way, like (say) tanh-sinh quadrature does. $\endgroup$ – user14717 Jun 27 '17 at 16:18
3
$\begingroup$

Use the identity \begin{align*} \int_{\mathbb{R}^{n}}\exp(-S(\mathbf{x})) \delta(f(\mathbf{x})) \, \mathrm{d}\mathbf{x} = \int_{f^{-1}(0) \subset \mathbb{R}^{n-1}} \frac{\exp(-S(\mathbf{x})) }{|\nabla f(\mathbf{x})|} \, \mathrm{d}\mathbf{x} \end{align*} and then use Monte-Carlo on the surface integral. The difficulty will be accurate determination of the surface $f^{-1}(0)$.

Note that naive application of Monte-Carlo integration to an integrand with a delta-function will never work. You attempting to randomly sample a set with measure zero, so unless some finite-precision arithmetic works in your favor, the integral will always evaluate to zero.

$\endgroup$
2
$\begingroup$

One thought: the Dirac Delta is technically defined as the limit of a sequence of functions (like box functions or Gaussians). Just take one of these functions for your Dirac Delta and perform MC on that. That is, use

$$ \delta(x) \approx \frac{1}{n\sqrt{\pi}} e^{-(x/n)^2} $$

as an approximation to the Dirac Delta, for some suitably large $n$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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