It is to use the certain kind of model with mixed linear and nonlinear state in current navigation. Its characteristic is that measure function is a nonlinear function of some states and state equation is linear or weak nonlinear system of equations. The problem can be solved through marginal particle filter with state decomposition. The problem is solved through state decomposition of Rao-Blackwellised theory, which nonlinear state is got by particle filter. On the basis particle of nonlinear state, marginalize out linear state, which solved analytically through Kalman filter. The article provides theory study and demonstrates the flow of algorithm. With the study of underwater gravity gradient aided navigation, its condition meets on the condition of this method. So give simulation application of it, and have a comparison with APO-PF in the same large initial position error, the result indicates that the algorithm has a high efficient effect on the underwater gravity gradient navigation.