Improved Sampling‐Importance Resampling and Reduced Bias Importance Sampling (Q4828215)

An improvement of the sampling-importance resampling (SIR) algorithm is proposed. It aims at drawing a random sample from a target distribution \(\pi(y)\propto h(y)q(y)\), where \(q\) is a probability density and \(h\) is a computable function. At the first step of the improved SIR without replacement a sample \(X=(X^1,\dots,X^n)\) is drawn from \(q\). Then \(Y^1\) is drawn from \(X\) with probability \(w(X^i)\propto h(X^i)/(\sum_{j=1}^n h(X^j)-h(X^i))\) and \(Y^k\), \(k=2,\dots,m\) are drawn from the rest of \(X\) with probabilities at the \(k\)-th step \[ w(X^i)\propto h(X^i)/ \left(\sum_{j=1}^n h(X^j)-\sum_{j=1}^{k-1} h(Y^j)-k h(X^i)\right). \] It is shown that if \(p^n(y)\) is the distribution of \(Y\), then \(\| p^n-\pi\|_{TV}\leq O(n^{-2})\) (\(m\) is fixed), where \(\|\cdot\|_{TV}\) is the total variance norm. The reduced bias importance sampling for the estimation of \(E_\pi g(X)\) is also considered.