Abstract. A nonlinear test model for filtering turbulent signals from partial observations of nonlinear slow-fast systems with multiple time scales is developed here. This model is a nonlinear stochastic real triad model with one slow mode, two fast modes, and catalytic nonlinear interaction of the fast modes depending on the slow mode. Despite the nonlinear and non-Gaussian features of the model, exact solution formulas are developed here for the mean and covariance. These formulas are utilized to develop a suite of statistically exact extended Kalman filters for the slow-fast system. Important practical issues such as filter performance with partial observations, which mix the slow and fast modes, model errors through linear filters for the fast modes, and the role of observation frequency and observational noise strength are assessed in unambiguous fashion in the test model by utilizing these exact nonlinear statistics.Key words. Nonlinear model, slow-fast system, extended Kalman filter AMS subject classifications. 60H10, 60G35
IntroductionMany contemporary problems in science and engineering involve large dimensional turbulent nonlinear systems with multiple time scales, i.e., slow-fast systems. The increasing need for real time predictions, for example, in extended range forecasting of weather and climate, drives the development of improved strategies for data assimilation or filtering [7,13,6,10,2,3,28,4,14]. Such filtering algorithms are based on generalizations of the classical Kalman filter [1,5]. Filtering combines partially observed features of the chaotic turbulent multiscale signal together with a dynamic model to obtain a statistical estimate for the state of the system. The dynamic models for the coupled atmosphere-ocean system are prototype examples of slow-fast systems where the slow modes are advective vortical modes and the fast modes are inertia-gravity waves [29,9,21]. Depending on the spatio-temporal scale, one might need only a statistical estimate of the slow modes, as on synoptic scales in the atmosphere [7] or both slow and fast modes such as for squall lines on mesoscales due to the impact of moist convection [17]. In either situation, the noisy partial observations of quantities such as temperature, pressure, and velocity necessarily mix both the slow and fast modes [7,6,21]. Furthermore, the dynamical models often suffer from significant model errors due to lack of resolution or inadequate parametrization of physical processes such as clouds, moisture, boundary layers, and topography.The goal of the present paper is to develop a simple three dimensional nonlinear test model with exactly solvable statistics for slow-fast systems in order to provide unambiguous guidelines for the difficult issues for filtering slow-fast systems from partial observations, as mentioned in the first paragraph. Here, we also study filter performance and model errors in this idealized low-dimensional setting for slow-fast filtering; this study parallels earlier works [11,26] using low dimensional nonlinear ...