We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Mathematics > Statistics Theory

Title: Grid Based Nonlinear Filtering Revisited: Recursive Estimation & Asymptotic Optimality

Abstract: We revisit the development of grid based recursive approximate filtering of general Markov processes in discrete time, partially observed in conditionally Gaussian noise. The grid based filters considered rely on two types of state quantization: The \textit{Markovian} type and the \textit{marginal} type. We propose a set of novel, relaxed sufficient conditions, ensuring strong and fully characterized pathwise convergence of these filters to the respective MMSE state estimator. In particular, for marginal state quantizations, we introduce the notion of \textit{conditional regularity of stochastic kernels}, which, to the best of our knowledge, constitutes the most relaxed condition proposed, under which asymptotic optimality of the respective grid based filters is guaranteed. Further, we extend our convergence results, including filtering of bounded and continuous functionals of the state, as well as recursive approximate state prediction. For both Markovian and marginal quantizations, the whole development of the respective grid based filters relies more on linear-algebraic techniques and less on measure theoretic arguments, making the presentation considerably shorter and technically simpler.
Comments: 38 pages. To appear in the IEEE Transactions on Signal Processing
Subjects: Statistics Theory (math.ST); Information Theory (cs.IT); Optimization and Control (math.OC); Methodology (stat.ME); Machine Learning (stat.ML)
DOI: 10.1109/TSP.2016.2557311
Cite as: arXiv:1604.02631 [math.ST]
  (or arXiv:1604.02631v1 [math.ST] for this version)

Submission history

From: Dionysios Kalogerias [view email]
[v1] Sun, 10 Apr 2016 03:06:23 GMT (143kb,D)

Link back to: arXiv, form interface, contact.