Guaranteed Conservative Fixed Width Confidence Intervals Via Monte Carlo Sampling
Monte Carlo methods are used to approximate the means, $μ$, of random variables $Y$, whose distributions are not known explicitly. The key idea is that the average of a random sample, $Y_1, ..., Y_n$, tends to $μ$ as $n$ tends to infinity. This article explores how one can reliably construct a confidence interval for $μ$ with a prescribed half-width (or error tolerance) $\varepsilon$. Our proposed two-stage algorithm assumes that the kurtosis of $Y$ does not exceed some user-specified bound. An initial independent and identically distributed (IID) sample is used to confidently estimate the variance of $Y$. A Berry-Esseen inequality then makes it possible to determine the size of the IID sample required to construct the desired confidence interval for $μ$. We discuss the important case where $Y=f(\vX)$ and $\vX$ is a random $d$-vector with probability density function $ρ$. In this case $μ$ can be interpreted as the integral $\int_{\reals^d} f(\vx) ρ(\vx) \dif \vx$, and the Monte Carlo method becomes a method for multidimensional cubature.