Improving the Normalized Importance Sampling Estimator


The normalized importance sampling estimator allows the target density f to be known only up to a multiplicative constant. We indicate how it can be derived by a delta method-based approximation of a Rao–Blackwellized acceptance rejection estimator. Using additional terms in the delta method then results on a new estimator that also only requires f to be known only up to a multiplicative constant. Numerical examples indicate that the new estimator usually outperforms the normalized importance sampling estimator in terms of mean square error.


Sheldon Ross
Publication date: 
October 1, 2012
Publication type: 
Journal Article
Probability in the Engineering and Informational Sciences, 26 (4), 2012.