Aviso: Se está a ler esta mensagem, provavelmente, o browser que utiliza não é compatível com os "standards" recomendados pela W3C. Sugerimos vivamente que actualize o seu browser para ter uma melhor experiência de utilização deste "website". Mais informações em

Warning: If you are reading this message, probably, your browser is not compliant with the standards recommended by the W3C. We suggest that you upgrade your browser to enjoy a better user experience of this website. More informations on

Instituto Superior de Economia e Gestão (ISEG)


Seminário CEMAPRE "Error Dynamics in Nonlinear Prediction Systems - an Analytical Approach to a Numerical Problem"

Em 10-02-2012 às 11:00 / Edifício Quelhas, sala UNICRE

Rui A. P. Perdigão - Instituto Dom Luiz, Universidade de Lisboa, Portugal

Abstract: Nonlinear prediction systems are known to suffer from fundamental uncertainties associated with their sensitivity to the initial conditions and with the inaccuracy in the model representation. In order to address this issue, a systematic formulation is derived for the dynamics of prediction errors in numerical models under the combined influence of initial-condition and model-related errors. Analytical tests are conducted for chaotic systems and their results are confronted with those from numerical experiments. In doing so, some generic features for the error dynamics are brought out and connected with intrinsic physical properties of the underlying system. Analytical and numerical results are seen to agree within the domain of validity of the formulation, the extension of which is currently in progress. The main advantages of the proposed formulation with respect to empirical numerical tests consist on three core aspects: 1) The formulation allows for a quick evaluation of the dynamics of prediction errors in an analytical fashion, without the need to always run the numerical model under analysis; 2) It reveals the role played by different error contributions and their relation to intrinsic system properties; 3) The formulation is model-independent, therefore it allows for general properties to be revealed that help understand better the dynamics of prediction errors.