I have the following question regarding the Gibbs sampler, although it might be considered a simple question on conditional probability.

For sake of simplicity, let us say we are trying to sample from a bivariate normal distribution: $$ [X,Y]^T \sim \left([0,0], \begin{bmatrix}1 & \rho \\ \rho & 1\end{bmatrix}\right) $$ To perform Gibbs sampling, we generate samples: $$ x_{n+1} \sim X|Y^{(n)} $$ $$ y_{n+1} \sim Y|X^{(n+1)} $$ and output samples $(x_n,y_n)$. Here the notation $X|Y$means in words "the conditional distribution of $X$ given $Y$.

Given this information, is it possible to calculate the "marginalized" conditional distribution: $$ X_{n+1}|X_n $$ directly? How would one do it? I assume this is a basic conditional probability fact unrelated to Gibbs sampling.


Here is an excerpt (p.340) from our book Monte Carlo Statistical Methods that gives an answer to this question:

Example 9.1. Normal bivariate Gibbs. For the special case of the bivariate normal density, \begin{equation} \tag{9.3} (X, Y) \sim {\cal N}_2 \left( 0, \left( \begin{array}{cc} 1 & \ \ \rho \\ \rho & \ \ 1 \end{array} \right) \right) \;, \end{equation} the Gibbs sampler is

Given $y_t$, generate \begin{eqnarray}\tag{9.4} X_{t+1} \mid y_t & \sim &{\cal N}(\rho y_t,\ 1 - \rho^2) \;,\\ Y_{t+1} \mid x_{t+1} & \sim &{\cal N}(\rho x_{t+1},\ 1- \rho^2). \nonumber \end{eqnarray}

The Gibbs sampler is obviously not necessary in this particular case, as iid copies of $(X,Y)$ can be easily generated using the Box-Muller algorithm (see Example 2.8). Note that the corresponding marginal Markov chain in $X$ is defined by the AR$(1)$ relation $$ X_{t+1} = \rho^2 X_t + \sigma\epsilon_t,\qquad \epsilon_t\sim\mathcal{N}(0,1)\,, $$ with $\sigma^2 = 1-\rho^2+\rho^2(1-\rho^2) = 1 - \rho^4$. As shown in Example 6.43,the stationary distribution of this chain is indeed the normal distribution $$\mathcal{N}\left(0,\frac{1-\rho^4}{1 - \rho^4}\right)$$

  • 1
    $\begingroup$ Thank you, I'll do the calculations in some detail and then accept the answer. $\endgroup$ – rubikscube09 May 30 '20 at 16:29
  • 1
    $\begingroup$ Yes, I realize now that one must use the tower property: $\mathbb{E}[[X_{n+1}|Y_n]|x_n] = \mathbb{E}[X_{n+1}|X_n]$ $\endgroup$ – rubikscube09 May 30 '20 at 17:16

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.