If $f_1,\ldots,f_k$ are known densities from which I can simulate, i.e., for which an algorithm is available. and if the product $$\prod_{i=1}^k f_i(x)^{\alpha_i}\qquad \alpha_1,\ldots,\alpha_k>0$$ is integrable, is there a generic approach to simulate from this product density using the simulators from the $f_i$'s?

  • 2
    $\begingroup$ Without additional assumptions, this seems unlikely. (Let $\alpha_i=1$ for simplicity. Let $\epsilon\gt 0$ be small. Suppose that associated with each $f_i$ is an interval $I_i$ on which $f_i\le 1$ and $\Pr_i(I_i)\gt 1-\epsilon$, outside of which $0\lt f_i\lt \epsilon$, and $I_i\cap I_j=\emptyset$ for $i\ne j$. Then the separate generators would almost always produce values in $I_i$, but the probability of $\prod f_i$ could be concentrated anywhere, seemingly unrelated to the $I_i$.) So, what else can you tell us about the $f_i$? $\endgroup$
    – whuber
    Commented Apr 10, 2016 at 19:42
  • 1
    $\begingroup$ (+10) Correct! Using a smaller $\alpha_i$ would however lead to flatten all elements and hence favour overlap of their effective supports... $\endgroup$
    – Xi'an
    Commented Apr 11, 2016 at 8:10
  • 1
    $\begingroup$ As whuber said tightness will be a problem, so I would take a transformation(OR preferential sampling) to cancel the tightness before generating random samples. There is one constructive approach I think I read a while ago. Sec 10.7 of link.springer.com/chapter/10.1007/978-1-4612-0209-7_10 Not sure if the discretization can also be applied here. $\endgroup$
    – Henry.L
    Commented Mar 14, 2017 at 20:09

1 Answer 1


Well, of course there's the acceptance-rejection algorithm, which I would implement for your example as:

  1. (Initialization) For each $i$, find $A_i = \sup_x \{\Pi_{j=1}^k f_j(x)^{\alpha_j}/f_i(x)\} $. Edit reflecting Xi'an's comment below: Select the distribution $f_i$ which corresponds to the smallest $A_i$.
  2. Generate $x$ from $f_i$.
  3. Calculate $\alpha = \Pi_{i=j}^k f_j(x)^{\alpha_j} / (A_if_i(x))$.
  4. Generate $u \sim U(0,1)$.
  5. If $u \leq \alpha$, return $x$, else go to 2.

Depending on the distributions, of course, you might have a very low acceptance rate. As it happens, the expected number of iterations is equal to the selected $A_i$ (assuming continuous distributions), so at least you are warned in advance.

  • 4
    $\begingroup$ (+1) Indeed a solution! Assuming the bounds $A_i$ do exist for all $i$'s. Or even some $i$'s. Comparing the $A_i$'s [assuming they are finite] may also help in selecting the most efficient $f_i$. $\endgroup$
    – Xi'an
    Commented Sep 24, 2017 at 13:16
  • 1
    $\begingroup$ I hadn't thought of it, but of course you're right, the $A_i$s themselves are very informative, as they are also equal to the expected number of iterations needed to actually generate the random number if you stick with one $i$ throughout. So you would want to pick the distribution with the smallest $A_i$ to use all the time. I'll edit the answer so your point doesn't get lost in comments. $\endgroup$
    – jbowman
    Commented Sep 24, 2017 at 15:45
  • $\begingroup$ That is, assuming all $f_i$'s are properly normalised [as to integrate to one], which is not necessarily a standard occurrence. $\endgroup$
    – Xi'an
    Commented Sep 24, 2017 at 16:57

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.