I came across this question in a book and having trouble understanding it.

Suppose $(X,Y)$ are continuous random vector with joint pdf $f(x,y)$ and support $\mathcal{X} \times \mathcal{Y}$. In particular suppose that the marginals $f_X(x)$ and $f_Y(y)$ have support $\mathcal{X}$ and $\mathcal{Y}$ respectively. I need to verify that $k(x,y|x^\prime, y^\prime) = f(x|y^\prime)f(y|x)$ is a density function.

I think this problem has to do with applying Fubini theorem but what confuses me are $x^\prime$ and $y^\prime$. There is no mention to what they mean.Should I consider them to be other random variables distinct from $x$ and $y$ but with supports in $\mathcal{X}$ and $\mathcal{Y}$? I am not sure of how to understand this problem. Any help would be appreciated.

  • 1
    $\begingroup$ There must be a typographical error in the crucial part of your question, because "$x^\prime$" appears nowhere on the right hand side of the defining equation for $k$. Could you please fix this? $\endgroup$ – whuber Jun 5 '16 at 22:37
  • 1
    $\begingroup$ Yes that is what I was wondering too but I think it is not a typographical error because, in the next part of the question we use this $k$ to show that $f(x,y) = \int_{\mathcal{X}} \int_{\mathcal{Y}}k(x,y|x^\prime,y^\prime) f(x^\prime, y^\prime) d y^\prime d x^\prime$ and I used the same function $k$ to show this holds. $\endgroup$ – user111092 Jun 5 '16 at 22:42
  • 1
    $\begingroup$ The very equation is inherently ambiguous: it doesn't even use the notation "$f$" correctly, because $f$ is a function of two variables! That ambiguity means we can't even deduce what the domains of $x,y,x^\prime,y^\prime$ are intended to be or which of them are intended to be fixed and the others to be conditioned on them. $\endgroup$ – whuber Jun 5 '16 at 22:47
  • $\begingroup$ Exactly my trouble. I was trying to make sense out of the problem but I guess it should have been more descriptive. I am guessing $x^\prime$ and $y^\prime$ are other random variables with supports in $\mathcal{X}$ and $\mathcal{Y}$. $\endgroup$ – user111092 Jun 5 '16 at 22:51
  • $\begingroup$ Could you give a reference to this book? $\endgroup$ – Xi'an Jun 6 '16 at 9:55

Apart from an abuse of notation ($f$ having multiple meanings), there is no true difficulty in defining $$ k(x,y|x^\prime, y^\prime) = f(x|y^\prime)f(y|x) $$ as a density on the pair $(X,Y)$ since

  1. The conditional density of $Y$ given $X$ is $f(y|x)$, which integrates to one;
  2. The marginal density of $X$ is $f(x|y')$, which also integrates to one.

What may sound confusing is the parametrisation in $(x',y')$ but this simply indicates that the distribution does not depend on $x'$. A standard occurrence of this setting is in Gibbs sampling or slice sampling (which is possibly where you found this description): when using two full conditionals for the Markov transition, one simulates [at iteration $t$]

  1. $X_{t+1}|X_t,Y_t\sim f(x|y_t)$
  2. $Y_{t+1}|X_t,Y_t,X_{t+1} \sim f(y|x_{t+1})$

or $$(X_{t+1},Y_{t+1})|(X_{t},Y_{t})\sim k(x,y|x_t,y_t)$$ and the move does not directly depend on the value of $X_t$.

Note that this is not incompatible with $$(X_{t+1},Y_{t+1})\sim f(x,y)$$ in the stationary regime of the Markov chain, since $f(x,y)$ is the stationary density while $k(x,y|x_t,y_t)$ is the conditional density at iteration $t+1$.


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.