Nonlinear data assimilation in highly nonlinear large-dimensional systems
|
Fri, 04/05/2012 14:30 |
Prof. Peter Jan van Leeuwen (University of Reading) |
Mathematical Geoscience Seminar  |
DH 3rd floor SR |
| Data assimilation in highly nonlinear and high dimensional systems is a hard problem. We do have efficient data-assimilation methods for high-dimensional weakly nonlinear systems, exploited in e.g. numerical weather forecasting. And we have good methods for low-dimensional (<5) nonlinear systems. The combination is more difficult, however. Recently our data-assimilation group managed to generate efficient particle filters that seem to scale almost perfectly with the dimension of the system, that is the number of particles (model runs) needed is independent of the system dimension. This will be demonstrated on the barotropic vorticity equations in the chaotic regime, exploring different observation strategies. The main question now is why these methods are so efficient. The performance seems to be independent of traditional measures of stability, such as the number of positive Lyaponov exponents or decorrelation times of the dynamics. Our latest progress in this area will be discussed, bringing in elements of extreme value statistics and the stability of the combined model/observation system. | |||