A utility based approach to information for stochastic differential equations (Q1312299): Difference between revisions

Suppose the observed process \(X^ T= \{X_ t\), \(0\leq t\leq T\}\) satisfies \[ dX_ t= S(\theta,t,X_ t)dt+ \sigma(t,X_ t)dB_ t, \] where \(\{B_ t\), \(t\geq 0\}\) is a Brownian motion, and \(\theta\) is the parameter of interest. Denote by \(p(\theta)\) and \(p(\theta| X^ T)\) the prior and posterior densities for the parameter, respectively. The likelihood function is \[ L_ T(\theta)= \exp \biggl\{ \int_ 0^ T \bigl[ S(\cdot,t,X_ t)/\sigma^ 2(t,X_ t)\bigr]dt- 2^{-1} \int_ 0^ T \bigl[ S(\cdot,t,X_ t)/\sigma(t,X_ t)\bigr]^ 2dt\biggr\} \] and \(i_ T(\theta)= \langle (d/d\theta)\log L_ T(\theta)\rangle\) is the observed Fisher information. The main result is that under some regularity conditions: \[ \lim_{T\to\infty} \bigl\{ E_ X T_{,\theta} [\log (p(\theta| x^ T)/p(\theta))]- E_ X T_{,\theta} [\log| i_ T(\theta)|^{1/2}]\bigr\}= H(p)- 2^{-1}\log(2\pi e), \] where \(H(p)=- \int p(\theta)\log p(\theta)d\theta\) is the entropy of the parameter. Applications to Brownian motion with drift, Ornstein-Uhlenbeck processes and Bessel processes are given. There are some misprints. Even the conclusions in Theorem 4.1 are missed.