Looking for a recursive formula to approximate variance of importance sampling estimator $Var_q\big[\delta_{IS}\big]\approx\sum_{i=1}^n\tilde w(X_i)^2\big[h(X_i)-\delta_{IS}\big]^2$. This is an approximation of $$Var\big[\delta_{IS}\big]=E_q\Bigg\{w(X)^2\Big[h(X)-E_q\big[h(X)\big]\Big]^2\Bigg\}$$

I'm using this to make running variance plot.


1 Answer 1



\begin{align} w_i &= w \left( X_i \right) \\ h_i &= h \left( X_i \right) \\ \mu^w_N &= \frac{1}{N} \sum_{i = 1}^N w_i \\ \mu^{wh}_N &= \frac{1}{N} \sum_{i = 1}^N w_i h_i \\ \tilde{\mu}^h_N &= \frac{\mu^{wh}_N}{\mu^w_N} \end{align}

so that $\tilde{\mu}^h_N$ is the standard SNIS estimator.

The standard estimator of the asymptotic variance for SNIS is then given by

\begin{align} V_N &= \frac{\frac{1}{N}\sum_{i = 1}^N \left\{ w_i^2 \cdot \left( h_i - \tilde{\mu}^h_N \right)^2 \right\} }{\left( \frac{1}{N} \sum_{i = 1}^N w_i \right)^2} \end{align}

Expand this to see that

\begin{align} V_N &= \frac{\left( \frac{1}{N} \sum_{i = 1}^N w_i^2 h_i^2 \right) - 2 \cdot \tilde{\mu}^h_N \cdot \left( \frac{1}{N} \sum_{i = 1}^N w_i^2 h_i \right) + \left( \tilde{\mu}^h_N \right)^2 \cdot \left( \frac{1}{N} \sum_{i = 1}^N w_i^2 \right)}{\left( \frac{1}{N} \sum_{i = 1}^N w_i^2 \right)} \\ &= \frac{\mu^{wwhh}_N - 2 \tilde{\mu}^h_N \cdot \mu^{wwh}_N + \left( \tilde{\mu}^h_N \right)^2 \cdot \mu^{ww}_N}{\left( \mu^{w}_N \right)^2} \\ &= \frac{\left( \mu^w_N \right)^2 \cdot \mu^{wwhh}_N - 2 \mu^w_N \cdot \mu^{wh}_N \cdot \mu^{wwh}_N + \left( \mu^{wh}_N \right)^2 \cdot \mu^{ww}_N}{ \left( \mu^w \right)^4} \end{align}


\begin{align} \mu^{ww}_N &= \frac{1}{N} \sum_{i = 1}^N w_i^2 \\ \mu^{wwh}_N &= \frac{1}{N} \sum_{i = 1}^N w_i^2 h_i \\ \mu^{wwhh}_N &= \frac{1}{N} \sum_{i = 1}^N w_i^2 h_i ^2. \end{align}

As such, the estimator $V_N$ can be computed as a function of the statistics $\mu^w_N, \mu^{ww}_N, \mu^{wh}_N, \mu^{wwh}_N, \mu^{wwhh}_N$. Each of these statistics is a sample average, which can easily be updated recursively.

A modification of the above procedure which may provide improved numerical stability would be to note that

\begin{align} \mu^{ww}_N &= \left( \mu^{w}_N \right)^2 + \frac{1}{N} \sum_{i = 1}^N \left( w_i - \mu^{w}_N \right)^2 \\ \mu^{wwh}_N &= \mu^{wwh}_N \cdot \mu^{wh}_N + \frac{1}{N} \sum_{i = 1}^N \left( w_i - \mu^{ww}_N \right) \left(w_i h_i - \mu^{wh}_N \right) \\ \mu^{wwhh}_N &= \left( \mu^{wh}_N \right)^2 + \frac{1}{N} \sum_{i = 1}^N \left( w_i h_i - \mu^{wh}_N \right)^2, \end{align}


\begin{align} \beta^w_N &= \frac{1}{N - 1} \sum_{i = 1}^N \left( w_i - \mu^{w}_N \right)^2 \\ \rho^{w, wh}_N &= \frac{1}{N - 1} \sum_{i = 1}^N \left( w_i - \mu^{w}_N \right) \left(w_i h_i - \mu^{wh}_N \right) \\ \beta^{wh}_N &= \frac{1}{N - 1} \sum_{i = 1}^N \left( w_i h_i - \mu^{wh}_N \right)^2, \end{align}

and use standard, stable, recursive update formulas for variances and covariances to update $\left(\beta^w_N, \rho^{w, wh}_N, \beta^{wh}_N \right)$ in addition to $\left( \mu^{w}_N, \mu^{wh}_N \right)$. These estimators are all unbiased for their corresponding ideal quantities and are otherwise well-behaved, and so may be useful for other diagnostic purposes.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.