Let $(E,\mathcal E,\lambda)$ be a measure space, $p:E\to[0,\infty)$ be $\mathcal E$-measurable with $$c:=\int p\:{\rm d}\lambda$$ and $$\mu:=\underbrace{\frac1cp}_{=:\:\tilde p}\lambda$$ denote the measure with density $\tilde p$ with respect to $\lambda$.

Now let $q:E:\to[0,\infty)$ be $\mathcal E$-measurable with $$\int q\:{\rm d}\lambda=1$$ and $$\left\{q=0\right\}\subseteq\left\{p=0\right\}\tag1$$ and $\nu:=q\lambda$. Suppose we're actually interested in the distribution $\mu$. Is there any reason why we might prefer to run the Metorpolis-Hastings with target distributon $\nu$ instead of $\mu$?

Clearly, if $(X_n)_{n\in\mathbb N_0}$ and $(Y_n)_{n\in\mathbb N}$ denote the processes generated by the Metropolis-Hastings algorithm with target distribution $\mu$ and $\nu$, respectively, then $$\frac cn\sum_{i=b}^{b+n-1}f(X_i)\xrightarrow{n\to\infty}c\int f\:{\rm d}\mu=\int pf\:{\rm d}\lambda\tag2$$ and $$\frac1n\sum_{i=b}^{b+n-1}\frac{pf}q(Y_i)\xrightarrow{n\to\infty}\int\frac{pf}q\:{\rm d}\nu=\int pf\:{\rm d}\lambda$$ $\lambda$-almost everywhere for all $\mathcal E$-\measurable $f:E\to\mathbb R$ with $pf\in L^1(\lambda)$ and $b\in\mathbb N_0$. So, both processes can be used to approximate the integral on the right-hand side of $(2)$ and $(3)$.

I could imagine that the convergence to equilibrium can be sped up by choosing a suitable $q$. But this is just a vague speculation. Would be great if someone could elaborate on that.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.