Hello, I am trying to understand which of the assumptions underlying AIC and MDL lead to the different penalty terms (k for AIC, k lg N for MDL), especially in the context of autoregressive modeling. AIC assumes that if T* and T are the ideal and estimated parameter vectors, then for large N, (T*-T)' J (T*-T) ~ Chi-square where J is the Fisher matrix. The expected value of the above is k and hence the penalty. My question is this: as N increases, doesn't the Fisher matrix change also (i.e. doesn't the likelihood become more sensitive to changes in T)? If so, does incorporation of this sensitivity change the AIC penalty term to something like k lg N? (my understanding of the Fisher matrix is a little shaky ...). thanks in advance, Gautam Vallabha Complex Systems & Brain Sciences Florida Atlantic University