I am trying to conceptualize the analytical relationship between the prior distribution and posterior distribution obtained by MCMC for Bayesian inference. Sorry for the non-rigorous notation but I hope it still communicates the idea.

This thesis by Zhou suggests that the posterior $\Pr(X^n) = \Pr(X^n | \textit{data})$ of parameter $X$ can be obtained through $n$ acceptance steps via the prior $\Pr(X^0)$ and product of transition probabilities $\Pr(X^t | X^{t-1}) = \Pr(X^t | X^{t-1}, \textit{data})$: \begin{equation} \Pr(X^n) = \Pr(X^0) \prod_{t=1}^n \Pr(X^t | X^{t-1}) \end{equation} ($\textit{data}$ is omitted to simplify notation.)

Asher et al. writes that the limiting distribution that approximates the posterior distribution is obtained by: \begin{equation} \Pr(X^n) = \lim_{n\to\infty} \Pr(X^n | X^0) \end{equation}

According to the Chapman-Kolmogorov equation, \begin{align} \Pr(X^n | X^0) = \int_{-\infty}^{\infty} \Pr(X^n | x) \Pr(x | X^0) dx \end{align}

My question is: I am not sure this relationship equivalent? \begin{align} \Pr(X^n | X^0) \stackrel{?}{=} \Pr(X^0) \prod_{t=1}^n \Pr(X^t | X^{t-1}) \end{align}

Is there another way to conceptualize the relationship between $\Pr(X^0)$ and $\Pr(X^n)$ through transition probabilities?

  • $\begingroup$ This is most confusing as you only mention a sequence $(X_0,X_1,\ldots,X_n)$ and no parameter or observation. In which sense is this Bayesian or related with MCMC? $\endgroup$ – Xi'an Jun 24 '18 at 5:40
  • $\begingroup$ Thanks for the comment - I've now stated that $X$ is in fact the parameter and the observation or data is implied but not included to simplify the notation. $\endgroup$ – hatmatrix Jun 24 '18 at 19:09
  • $\begingroup$ Still confused by the fact that the parameter is of increasing size. What is the meaning of $n$? $\endgroup$ – Xi'an Jun 25 '18 at 5:20
  • $\begingroup$ $n$ is the index of iteration - while it looks like an exponent I have seen this notation to describe steps in MCMC. $\endgroup$ – hatmatrix Jun 25 '18 at 9:03

The equation \begin{equation} \Pr(X^n) = \Pr(X^0) \prod_{t=1}^n \Pr(X^t | X^{t-1}) \end{equation} is incorrect, it should be \begin{equation} \Pr(X^n) = \int \Pr(x^0) \prod_{t=1}^{n-1} \Pr(x^t | x^{t-1})\Pr(X^n | x^{n-1}) \,\text{d}x^0\cdots\,\text{d}x^{n-1} \end{equation}or \begin{equation} \Pr(X^n|X^0) = \int \prod_{t=1}^{n-1} \Pr(x^t | x^{t-1})\Pr(X^n | x^{n-1}) \,\text{d}x^1\cdots\,\text{d}x^{n-1} \end{equation}

  • 2
    $\begingroup$ This clears it up! Thanks - so basically you have to integrate between all states x$^1$ to x$^{n-1}$. So is it correct that the product should go from $t=1$ to $t=n-1$ rather than $t=1$ to $t=n$? $\endgroup$ – hatmatrix Jun 26 '18 at 12:37
  • 1
    $\begingroup$ Sorry, typo corrected! $\endgroup$ – Xi'an Jun 26 '18 at 15:10

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.