Tuan Anh Le

Effective Sample Size

14 September 2017

Using the notation from the note on importance sampling, we investigate effective sample size (ESS), , which is used to assess the quality of an importance sampler estimator: \begin{align} N_e &:= \frac{1}{\sum_{n = 1}^N w_n^2}, \end{align} where are the normalized weights.

Recall that variance of a Monte Carlo estimator, , is . Intuitively, should provide the following approximation: \begin{align} \Var\left[\tilde I_N^{\text{IS}}\right] \approx \frac{\Var[f(X)]}{N_e}, \end{align} i.e., it approximates the variance of the importance sampling estimator by the variance of the Monte Carlo estimator with direct samples from .

Derivation 1

(From Art Owen’s book chapter). Making an approximation that the weights are not random variables but fixed values, i.e. are the unnormalized weights and are the normalized weights. \begin{align} \Var\left[\tilde I_N^{\text{IS}}\right] &\approx \Var\left[\frac{\sum_{n = 1}^N \tilde w_n f(X_n)}{\sum_{n = 1}^N \tilde w_n}\right] \\ &= \Var\left[\sum_{n = 1}^N w_n f(X_n)\right] \\ &= \sum_{n = 1}^N w_n^2 \Var\left[f(X_n)\right]. \end{align} Rearranging to equal , we obtain \begin{align} N_e &= \frac{1}{\sum_{n = 1}^N w_n^2}. \end{align}

Derivation 2

(From Sebastian Nowozin’s blogpost).

Properties of ESS

High ESS, bad performance

\begin{align} p(x) &= 0.5 * \mathrm{Normal}(x; -10, 1) + 0.5 * \mathrm{Normal}(x; 10, 1) \\ q(x) &= \mathrm{Normal}(x; 10, 1) \\ f(x) &= \begin{cases} 1 & \text{ if } x < 0 \\ 0 & \text{ otherwise.} \end{cases} \end{align}

Low ESS, good performance

(From Exercise 9.11 from Art Owen’s book chapter).