Suppose, you want to calculate

begin{equation*}

theta = int_{0}^{1} h(x)dx

end{equation*}

where $h$ is some complicated but continuous function (though evaluating $h$ at a particular $x$ is not hard). For example, one such function might be

begin{equation*}

h(x) = exp(-sqrt{x}) | sin(x^4log(x))|

end{equation*}

(i) Suppose you can draw random observations from $g$, where $g$ is a positive density on $(0, 1)$. Let $X_1, . . . , X_n$ be i.i.d. from $g$. Consider the estimator of $theta$ given by

begin{equation*}

hat{theta_n} = frac{1}{n} sum_{i=1}^{n} frac{h(X_i)}{g(X_i)}

end{equation*}

(For example, if $g$ is the uniform density, you are just averaging the function over the data points.) Compute $E(hat{theta_n})$.

indent We calculate $E(hat{theta_n})$ as follows:

begin{equation*}

begin{aligned}

& E(hat{theta_n}) = \

& E(frac{1}{n} sum_{i=1}^{n} frac{h(x_i)}{g(x_i)}) = \

& frac{1}{n}int_{0}^{1} frac{h(x)}{g(x)} g(x) dx = \

& frac{1}{n}int_{0}^{1} h(x) frac{g(x)}{g(x)} dx = \

& frac{1}{n}int_{0}^{1} h(x) dx = frac{theta}{n} \

end{aligned}

end{equation*}

(ii) Assume g is the uniform density. Compute $Var(hat{theta_n})$ and give a good upper bound to this variance if you know that $0 geq h(x) geq 1$ for all $x$.

indent Using expectation we break apart the variance as follows:

begin{equation*}

begin{aligned}

& Var(hat{theta_n}) = \

& E(hat{theta_n}^2) – (E(theta_n)^2)= \

& E left(left( frac{1}{n} sum_{i=1}^{n} frac{h(x_i)}{g(x_i)} right)^2 right) – left(E left( frac{1}{n} sum_{i=1}^{n} frac{h(x_i)}{g(x_i)} right) right)^2 = \

& E left(frac{1}{n^2} left( sum_{i=1}^{n} frac{h(x_i)}{g(x_i)} right)^2 right) – left( frac{theta}{n} right)^2 = \

& frac{1}{n^2} left( int_{0}^{1} left(frac{h(x)}{g(x)}right)^2 g(x) right) – left( frac{theta}{n} right)^2 = \

& …\

end{aligned}

end{equation*}

Is this last step correct? Is my next step to multiply out the g(x) terms so one cancels? When I do that I get stuck. Would appreciate a push if anyone has one.

(iii) Assume g is the uniform density. Using Chebychevâ€™s inequality, how large must n be to guarantee that the absolute difference $|hat{theta_n} – theta|$ is no bigger than 0.05 with probability at least 95%.

(iv) If you could choose $g$, what choice would minimize $Var(hat{theta_n})$.