A method to reduce noise in experimental data with nonlinear time evolution is
presented. A locally linear model is used to obtain a trajectory consistent
with the dynamics as well as with the measured data. In contrast to previous
methods where the fit to the dynamics and the cleaning are done in seperate
steps, it is an (iterated) one step procedure. This is made possible by using
data both from the past and the future in the locally linear model. The method
is applied to both artificial and real data. Among others, it it leads to a
significant improvement of corelation dimension estimates.