This title certainly needs some clarification: I need to compute parameters $a_i$ for a helper function $f(\vec{a};k)$ (for grid interpolation) which is fitted to a number of values $y_k$ which are subject to evolution dictated by a system of ODE's $d y_k/d t = F_k(t,\vec{y})$. The Levenberg-Marquardt fit of the values $y_k$ to the function $f(\vec{a};k)$ is repeated in each time step $t_j$. The precise values found for the parameters $a_{i,j}$, which are varied in this fit, does not matter, but I found that in the presence of adaptive step-size control for the ODE this simple algorithm can lead to instabilities as it influences the value of the rhs-functions $F_k(t,\vec{y})$ at higher order. For the present purpose I can assume that the evolution of the parameters is smooth plus some noise $a_{i,j} = A_i(t_j)+\delta_{i,j}$ and the model for the interpolating functions $f(\vec{a};k)$ does not matter. My idea to fix the stability issues is therefore to make the parameters $a_i$ subject to some smoothening procedure.

The concept for this procedure in turn is to estimate the new values $a_{i,J}$ based on the previous ones $a_{i,j}$, $j<J$ and the new $\chi^2$ fit by adding to the function to be minimized in the fit a contribution proportional to $\sum_l \sigma_l (A_l(t_J)-a_{l,J})^2$. Here $\sigma_l$ are positive constant coefficients and $A_l(t_J)$ are the values suggested by the trend of the previous ones (only loosely related to $A_l(t_J)$ above). The obtained $a_{l,J}$ would automatically be consistent with constraints in the fit. (Note that this could be different if I first fit and later smooth) The question is which method to use to predict (to extrapolate) the values $A_l(t_J)$. I wanted to avoid using a simple polynomial or spline fit, because I definitely need to avoid overshooting. On the other hand a second $\chi^2$ fit of a model function seems to overdo it. Also due to the fact that the previous values, $a_{i,j}$, $j<J$, where themselves found based on the combined fit and I am not sure that this does not introduce any bias. What do you think? I also wondered if I could miss use a Sawatzki-Golay smoothing filter (assuming as an approximation that $t_{i+1}-t_i=const$ to avoid the overhead of a full $\chi^2$ fit. But, can it be used for extrapolation?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.