Let's consider a sampling algorithm which is sampling from the posterior distribution $P(X)$ of a random variable $X$. At a given iteration $t+1$ it proposes a new value $(X_{t+1})$ for $X$ using a symmetric proposal distribution $Q(X_{t+1}|X_t)$ around the old value $X_t$. Due to the symmetry of the proposal distribution, the standard Metropolis Hastings (MH) acceptance rate is as follows: accept $X_{t+1}$ with probability $\alpha=\min(1,P(X_{t+1})/P(X_t))$. Somewhere, I saw a slightly different acceptance rate, it says:

  • if $P(X_{t+1})>P(X_t)$ accept $X_{t+1}$ (This part is same as the MH algorithm)
  • else sample from the distribution $P=[P(X_t); P(X_{t+1})]/(P(X_t)+P(X_{t+1}))]$ (In this case the sampler will choose $X_{t+1}$ or $X_t$ depending on their posterior probabilities)

I was wondering whether the later is a valid MCMC sampler.

  • $\begingroup$ Would you mind to cite the source of the second approach? $\endgroup$ Feb 7, 2017 at 10:48
  • $\begingroup$ Why can't you try to prove that the distribution $P$ is stationary with this transition? $\endgroup$
    – Xi'an
    Feb 14, 2017 at 5:56

1 Answer 1


This alternative version sounds like Barker's (1965) and Hastings' (1970) versions when the acceptance probability is $$\alpha(x,x^\prime) = s(x,x^\prime) \left(1+\dfrac{p(x) q(x^\prime|x)}{p(x^\prime) q(x|x^\prime)}\right)^{-1}$$with $s(x,y)$ is a symmetric function keeping the above between 0 and 1 except that there should be no sure acceptance first, i.e., when $p(x^\prime)>p(x)$.

For instance, here is an illustration of why the proposal in the question does not work: enter image description here

This histogram represents the result of 10⁵ simulations, against the intended standard Normal target, using another standard Normal random walk as proposal.

While, if I remove the first step in the "alternative" algorithm, Barker's formula leads to the proper fit: enter image description here


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.