Data assimilation into nonlinear stochastic models

With very few exceptions, data assimilation methods which have been used or proposed foruse with ocean models have been based on some assumption of linearity or near-linearity. Thegreat majority of these schemes have at their root some least-squares assumption. While onecan always perform least-squares analysis on any problem, direct application of least squaresmay not yield satisfactory results in cases in which the underlying distributions are significantlynon-Gaussian. In many cases in which the behavior of the system is governed by intrinsicallynonlinear dynamics, distributions of solutions which are initially Gaussian will not remain soas the system evolves. The presence of noise is an additional and inevitable complicating factor.Besides the imperfections in our models which result from physical or computational simplifyingassumptions, there is uncertainty in forcing fields such as wind stress and heat flux which willremain with us for the foreseeable future. The real world is a noisy place, and the effects ofnoise upon highly nonlinear systems can be complex. We therefore consider the problem ofdata assimilation into systems modeled as nonlinear stochastic differential equations. When themodels are described in this way, the general assimilation problem becomes that of estimatingthe probability density function of the system conditioned on the observations. The quantitywe choose as the solution to the problem can be a mean, a median, a mode, or some otherstatistic. In the fully general formulation, no assumptions about moments or near-linearity arerequired. We present a series of simulation experiments in which we demonstrate assimilationof data into simple nonlinear models in which least-squares methods such as the (Extended)Kalman filter or the weak-constraint variational methods will not perform well. We illustratethe basic method with three examples: a simple one-dimensional nonlinear stochastic differentialequation, the well known three-dimensional Lorenz model and a nonlinear quasigeostrophicchannel model. Comparisons to the extended Kalman filter and an extension to the extendedKalman filter are presented. DOI: 10.1034/j.1600-0870.1999.t01-2-00002.x