I'm estimating a finite number ($\approx1e5$) of integrals $\lambda g$ using the Metropolis-Hastings algorithm with target distribution $\mu=\frac{p\lambda}c$ (where $c:=\lambda p\in(0,\infty)$) and the importance sampling estimator.

In each iteration, given the current state $x$, I need to compute the quantity $$\int\lambda({\rm d}y)\frac{\left|1_H(y)p(y)-\frac{p(y)}c\lambda(1_Hp)\right|^2}{r(x,y)}\tag1,$$ where $r(x,\;\cdot\;)$ is a probability density with respect to $\lambda$ and $H\in\mathcal E$ with $\lambda(H)\in(0,\infty)$.

I'm able to sample from $r(x,\;\cdot\;)\lambda$ and hence can estimate $(1)$ by ordinary Monte Carlo integration. However, since I need to compute $(1)$ for $\approx1e5$ different $H$ in each iteration of the Metropolis-Hastings algorithm, this approach is extremely slow (practically unusable).

Is there a better approach for a problem of this kind? If not: One of the biggest problems with the Monte Carlo approach is that I have no clue which sample size I should choose for my estimate of $(1)$. Is there a good rule which tells me that enough samples have been taken?


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.