Minimal Entropy Approximations and Optimal Algorithms

From MaRDI portal





Partical filters are a class of recursive algorithms which approximate the filtering conditional distributions by using discrete random measures of the form \(\sum_i{}\alpha_i\delta_{V_i}\), which is the empirical distributions associated for a set of random particles of mass \({\alpha_1},{\alpha_2},\dots \), located at \(V_1,V_2,\dots \). In this paper, the signal \(X\) is a Markov chain, and with the observation \(Y\), the process \((X,Y)\) is also a Markov chain. Let the filtering conditional distribution at time \(t\) be \({\bar{\pi_t}}\) and \(\bar{\mathcal{A}}_t^{\bar{\pi}}\) be the class of empirical random measures represented as \( \widetilde{a}_t =\frac {1}{n} \sum_{k=1}^{m_t}{\delta_{V_K}}\) with \(E(m_t)=n, E(\widetilde{a}_t)={\bar{\pi_t}}\) and \(\widetilde{a}_t\) satisfies \(E[\widetilde{a}_t|\widetilde{a}_0,\dots ,\widetilde{a}_{t-1}]=\widetilde{a}_{t-1}P^t\), where \(P\) is the transition matrix of the Markov chain \(X\). By using a tree based branching algorithm constructed by the authors, it is proven that a series of approximated random particle measure \(\widetilde{\pi}_t\) in \(\bar{\mathcal{A}}_t^{\bar{\pi}}\) can be constructed recursively such that \(\widetilde{\pi}_t\) minimizes the relative entropy \(H(\widetilde{\pi}_t|\bar{\pi}_t)\) for all \(t\geq 0\).











This page was built for publication: Minimal Entropy Approximations and Optimal Algorithms

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4796472)