Efficient particle methods for residual generation in partially observed SDEs

摘要

The problem of detecting a change in the drift coefficient of a partially observed stochastic differential equation is addressed. The score function, evaluated at the nominal value, is used as the residual, and only the problem of residual generation is considered. In the special case where the drift coefficient depends on the parameter only in directions that are affected by nondegenerate noise, an efficient numerical approximation of the residual is proposed, using particle filters.