Quadratic Approximation for Log-Likelihood Ratio Processes

P. 179–215.

Alexander Gushchin, Valkeila E.

We consider a sequence of general filtered statistical models with a finite-dimensional parameter. It is tacitly assumed that a proper rescaling of the parameter space is already done (so we deal with a local parameter) and also time rescaling is done if necessary. Our first and main purpose is to give sufficient conditions for the existence of certain uniform in time linear–quadratic approximations of log-likelihood ratio processes. Second, we prove general theorems establishing LAN, LAMN and LAQ properties for these models based on these linear–quadratic approximations. Our third purpose is to prove three theorems related to the necessity of the conditions in our main result. These theorems assert that these conditions are necessarily satisfied if (1) an approximation of a much more general form exists and a (necessary) condition of asymptotic negligibility of jumps of likelihood ratio processes holds, or (2) we have LAN property at every moment of time and the limiting models are continuous in time, or (3) we have LAN property, Hellinger processes are asymptotically degenerate at the terminal times, and the condition of asymptotic negligibility of jumps holds.

Keywords: limit theorems, local asymptotic mixed normality, local asymptotic normality, local asymptotic quadraticity, Hellinger process, contiguity, filtered statistical experiment, likelihood ratio process

Publication based on the results of: Вероятностные и статистические методы анализа сложных моделей, задаваемых стохастическими дифференциальными и разностными уравнениями (2017)

In book

Modern problems of stochastic analysis and statistics - Selected contributions in honor of Valentin Konakov

Heidelberg: Springer, 2017.