Importance sampling: Why does dividing by average of likelihood ratios also produce an estimator??

17 Views Asked by At

Please read this short snippet on importance sampling, and explain to me why both $F_{3n}$ and $F_{4n}$ are estimators of $F(x)$. I get the first two estimators, but why the last 2?

The context is that we estimate $F(x)$ for some distribution by instead simulating $x_1, ..., x_n$ from a distribution $P'$ with respect to which $F$ is absolutely continuous with density $p(x)$.

enter image description here

1

There are 1 best solutions below

0
On

In the last two estimators the denominator is close to 1. Here's why: By construction, the connection between $F$ and the density $p$ is $$ F(x)=\int_{-\infty}^x p(t)dP'(t). $$ Now if $X_i$ has distribution $P'$ then the expectation of $L_i:=p(X_i)$ is $$ \int p(t)dP'(t) = F(\infty)=1. $$ So the quantity $\frac1n\sum L_i$ is equal to one on average.