2
$\begingroup$

Be $f(x)$ a function. Suppose that $f(x)$ integrates to a finite value $k$:
$$\int_{-\infty}^{\infty}f(x)dx=k$$
The normalization constant of $f(x)$ is $1/k$. Monte Carlo integration can give an estimate $\hat{k}$ of $k$, so we can estimate the normalization constant simply by taking the reciprocal of this estimate ($1/\hat{k}$). According to the CLT, the Monte Carlo estimator has an asymptotic normal distribution, so a 95% confidence interval for $k$ is simply: $$[\hat{k}-1.96*SE,\hat{k}+1.96*SE]$$ The question is: how can I calculate a confidence interval for the normalization constant? Can I just calculate the reciprocal of the previous confidence interval?

$\endgroup$
2
$\begingroup$

Yes: the defining property of a 95% confidence interval $[L, U]$ for an unknown parameter $k$ is that $P(L < k < U) = 95\%$. Since $L < k < U \iff 1/U < 1/k < 1/L$, you are guaranteed that $P(1/U < 1/k < 1/L) = 95\%$, so $[1/U, 1/L]$ is a 95% CI for $1/k$.

As an alternative, you could also use the Delta Method, but I would only do this if I thought it would improve the asymptotic approximation (make the CLT "kick in faster").

$\endgroup$

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.