I have a question regarding the support of an importance sampling distribution with respect to the support of the original distribution function. I was reading that the support of the importance sampling distribution (say the importance distribution is denoted q(x), say the set Q contains all value of x which satisfies q(x)>0 ) has to cover the support of the original distribution (say the distribution denoted as p(x) and let D denoted the set of x which satisfies p(x)>0 ), so it means we have $$D \subset Q$$.

The goal is to estimate some integral say:


by using this instead:

$$\int_{Q} f(x) \frac{p(x)}{q(x)} q(x) dx $$

My question is that the author is saying q(x) doesn't have to be entirely bigger than zero as long as it is bigger than zero when $f(x)p(x)\neq 0$, i.e. $ x \in{Q}$ whenever $f(x)p(x) \neq 0$.

And next it goes about saying this:

$E_{q} \Bigg(\frac{f(X)p(x)}{q(x)} \Bigg)$ = $\int_{Q} \frac{f(x)p(x)}{q(x)} q(x) dx$ =
$\int_{D} f(x)p(x) dx$ + $\int_{Q \cap D^c} f(x)p(x) dx-\int_{D \cap Q^c} f(x)p(x) dx $

My question has to do with the very last line here, I don't really get why the expression on the very last line here would equal this:

$\int_{Q} \frac{f(x)p(x)}{q(x)} q(x) dx$.

It seems it is saying the set $Q$ is a union of the two sets: $D$ and $Q\cap D^c$. And $D\cap Q^c$ is the intersection of $D$ and $Q\cap D^c$.

My question is, is my understanding of the union and intersection correct regarding the above?

Because it seems it is based on the formula of:

$P(A\cup B) = P(A) + P(B) - P(A \cap B)$

If it is indeed based on the above simple "set union operation", then my question is how does the terms:

$\int_{D} f(x)p(x) dx$ + $\int_{Q \cap D^c} f(x)p(x) dx-\int_{D \cap Q^c} f(x)p(x) dx $

equal $\int_{Q} \frac{f(x)p(x)}{q(x)} q(x) dx$.

P.S. Hi Xian, yes it is here: http://statweb.stanford.edu/~owen/mc/Ch-var-is.pdf

It is just some notes I read from the pdf file.

  • $\begingroup$ Yes Xian, the link is there now. Thank you very much again for your clear explanation. $\endgroup$ – john_w Nov 25 '16 at 16:52
  • $\begingroup$ Most useful for readers, thank you!, especially since they are lecture notes by Art Owen, a major actor in the field of Monte Carlo methods. $\endgroup$ – Xi'an Nov 25 '16 at 17:55

This can be easily resolved via the use of (set) indicator functions:

My question is that the author is saying $q(x)$ doesn't have to be entirely bigger than zero as long as it is bigger than zero when $f(x)p(x)≠0$, i.e. $x∈Q$ whenever $f(x)p(x)≠0$.

Indeed, the only constraint for the identity to hold is that the support $Q$ of $q$ contains the support of the function $x \to f(x)p(x)$, because, if $H$ denotes this support$$\int_D f(x)p(x) dx=\int_D f(x)p(x)\mathbb{I}_{H}(x) dx=\int_{D\cap H} f(x)p(x) dx=\int_H f(x)p(x) dx$$

As for the second part, if $D\subset Q$, \begin{align} \int_{Q} f(x)p(x) dx&=\int_{Q} \{\mathbb{I}_D(x)+\mathbb{I}_{D^c}(x)\} f(x)p(x) dx\\ &=\int_{Q \cap D} f(x)p(x) dx+\int_{Q \cap D^c} f(x)p(x) dx \end{align} but otherwise, since $$\mathbb{I}_{Q\cup D}=\mathbb{I}_Q+\mathbb{I}_{D\cap Q^c}=\mathbb{I}_D+\mathbb{I}_{Q\cap D^c}$$ we have $$\mathbb{I}_Q=\mathbb{I}_D+\mathbb{I}_{Q\cap D^c}-\mathbb{I}_{D\cap Q^c}$$

| cite | improve this answer | |
  • $\begingroup$ Oh, Thanks a lot Xi'an. It is so clear I think. Also sorry I just have one more question. The author says that there is no need to worry about sampling from Q when q(x) =0 (i.e. when the denominator could be zero , q(x) = 0), because the author says "we will never see one". I am just wondering why is it true that when sampling from the distribution Q, we will never seen a sample x such that q(x)=0? $\endgroup$ – john_w Nov 25 '16 at 12:32
  • $\begingroup$ With probability 1, a realisation of$X$ from the density q belongs to the support of $q$, $\{x;\,q(x)>0\}$. Hence there is a zero probability to obtain a realisation outside this support and $q(X)>0$ with probability one. $\endgroup$ – Xi'an Nov 25 '16 at 12:40
  • 1
    $\begingroup$ Hi Xian, I have attached the link to the file that contains the materials in the edited question above now. $\endgroup$ – john_w Nov 25 '16 at 16:49

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.