1

Consider a random variable $X\sim f(x)$, such that $$ f(x)=\frac{1}{c}\times K(x)\propto K(x), $$ where c: normalizing constant, K(x): the kernel of the distribution (ie the part which involves $x$). $f$ is unknown and complicated, in the sense that it does not resemble any known distribution.

The M-H algorithm is designed to simulate from $f$ based only on the kernel $K(x)$, which makes its application rather complicated. I wonder how easy the M-H algorithm would become if we know the normalizing constant $c$?

Xi'an
  • 90,397
  • 9
  • 157
  • 575
Ludwig
  • 51
  • 3
  • This does not bring any information on how to simulate $f$. – Xi'an Dec 20 '20 at 14:15
  • Thanks for your comment. Well, it depends on the actual kernel. For instance, provided I know $c$, I could use inverse transform sampling (https://en.wikipedia.org/wiki/Inverse_transform_sampling) instead of M-H. This can be achieved by some numerical integration method. But if I still need to use MH, how that information would help me? that was my question. – Ludwig Dec 20 '20 at 14:24
  • 1
    However, the normalization constant will not cancel out on the M-H? – Fiodor1234 Dec 20 '20 at 14:30
  • 2
    Since the question is about M-H, my answer remains that knowing the constant $c$ is not helping. – Xi'an Dec 20 '20 at 14:30
  • 1
    Fiodor, I am afraid so, so we shall conclude that MH is not the best choice in this case, since it does not make any difference to know the constant. Thanks all. – Ludwig Dec 20 '20 at 14:39
  • See also the answers to https://stats.stackexchange.com/a/307940/7224 – Xi'an Dec 20 '20 at 15:00
  • A side usage of $c$ would be to exploit it as a control variate, i.e., a function whose expectation is known, but this does not help in designing a better Metropolis-Hastings algorithm. – Xi'an Dec 21 '20 at 09:34

0 Answers0