0
$\begingroup$

Similar question to posted here: Metropolis-Hastings using log of the density however my question is around sampling a random number from a uniform distribution. I am following the steps outlined in Murphy: Machine Learning A Probabilistic Perspective on page 850 for Metropolis Hastings algorithm. The difference is I am working with log probability densities as they are large and negative, so exponentiating them effecitvely results in 0.

enter image description here

I have started by taking the log of step 5: $$ log(\alpha) = log(P(x')) - log(P(x)) $$ Assuming q is symmetric which it is in my case. Then to compute $r$, I am taking the lesser of $log(1)$ and $log(\alpha)$, so essentially if $\alpha$ is negative, I am assigning it to $r$, otherwise I am assigning $r$ as 0. Now for the issue: what do I do at step 6? The link above doesn't describe this. Obviously $r$ is either 0 or negative, and so it would always be less than $u$ and hence we always step where we are. Also, $r$ isn't bounded below - $\alpha$ could be very negative (though unlikely), so it is hard to redefine a uniform distribution range to sample from. Any ideas how to proceed from here?

$\endgroup$
5
  • 3
    $\begingroup$ I think that if you took $log(\alpha)$ then you will have to take $log(u)$ to preserve inequality. $\endgroup$
    – jassis
    Feb 13 at 1:53
  • $\begingroup$ So draw a random number from 0 to 1, and then take the log of it? $\endgroup$
    – spacexyz
    Feb 13 at 1:54
  • $\begingroup$ Precisely! I didn't get to the point, but as I like samplers, I found this post umbertopicchini.wordpress.com/2017/12/18/… $\endgroup$
    – jassis
    Feb 13 at 2:14
  • $\begingroup$ @jassis Thanks for the help - I have managed to get it to work. Out of interest, I have written the code two ways. In one instance I exponentiate $log(\alpha)$ and then continue the method in the image in the OP, in a second method I do the whole method in log format, doing as you said above and taking the log of $u$. Out of curiosity, do you know if the two methods should be equivalent? Or if one would be better than the other? I suspect there may be a difference when comparing $u$ to an exponentiated $log(\alpha)$ vs comparing $log(\alpha)$ to $log(u)$ due to different scales? $\endgroup$
    – spacexyz
    Feb 13 at 3:27
  • 1
    $\begingroup$ Both methods give exactly the same answer, they are more than equivalent. Note also that the min step is unnecessary for the comparison. $\endgroup$
    – Xi'an
    Feb 13 at 9:25

1 Answer 1

0
$\begingroup$

In log scale, here is what your pseudocode should read from step 5 onwards:

5. Compute acceptance probability

$$\log(\alpha) = \log \frac{p(x')}{p(x)} = \log p(x')- \log p(x).$$

Compute

$$\log(r) =\min(0, \log(\alpha))$$

6. Sample $u \sim U(0,1)$

7. Set new sample to

$$x^{s+1} = \begin{cases} x' \quad \text{if} \quad \log(u) < \log(r) \\ x^s \quad \text{if} \quad \log(u) \geq \log(r)\end{cases}$$


As you've correctly noted, because $q$ is symmetric, we have that $q(x | x') = q(x' | x)$ and so these cancel in your expression for $\log (\alpha)$.

For $\log(r)$, note that $\log(r) = \log \min(1, \alpha) = \min(\log(1), \log(\alpha)).$

Now that you have $\log(r)$ you just need to compare it to $\log(u)$ as we can log both sides of the inequalities in the decision rule.

$\endgroup$

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.