A general theory of particle filters in hidden Markov models and some applications (Q2443206): Difference between revisions

Let \({\text{X}} = \left\{ {{X_t}: t \geqslant 1} \right\}\) be a Markov chain and let \({\text{Y}} = \left\{ {{Y_t}: t \geqslant 1} \right\}\) be conditionally independent given X, such that \({X_t} \sim {p_t}\left( { \cdot \left| {{X_{t - 1}}} \right.} \right),{Y_t} \sim {g_t}\left( { \cdot \left| {{X_t}} \right.} \right)\), where \({p_t}\) and \({g_t}\) are density functions with respect to some measures \({\nu _X}\) and \({\nu _Y}\), and \({p_1}\left( { \cdot \left| {{X_0}} \right.} \right)\) denotes the initial density \({p_1}\left( \cdot \right)\) of \({X_1}\). Letting \({{\text{X}}_t} = ({X_1},\dots,{X_t})\), \({{\text{Y}}_t} = ({Y_1},\dots,{Y_t})\), and \(\psi \) be a measurable real-valued function of \({{\text{X}}_t}\), the paper considers estimation of \({\psi _T}: = \operatorname{E}[\psi ({{\text{X}}_T}){\left| {\text{Y}} \right._T}]\). The density function of \({{\text{X}}_T}\) conditional on \({{\text{Y}}_T}\) is often difficult to sample from and the normalizing constant is also to difficult to compute. Particle filters that use sequential Monte Carlo (SMC) methods involving importance sampling and resampling have been developed to circumvent this difficulty. The paper provides a comprehensive theory of the SMC estimate \({\hat \psi _T}\), which includes asymptotic normality and consistent standard error estimation.