If there is a good reason to assume that the relation is
fulfilled by the experimental data in good approximation (say, within 5%) for
some unknown f and that f is smooth, predictions can be improved by
fitting local linear models. They can be considered as the local Taylor
expansion of the unknown f, and are easily determined by minimizing
with respect to and , where is the -neighborhood of , excluding itself, as before. Then, the prediction is . The minimization problem can be solved through a set of coupled linear equations, a standard linear algebra problem. This scheme is implemented in onestep. For moderate noise levels and time series lengths this can give a reasonable improvement over zeroth and predict. Moreover, as discussed in Sec., these linear maps are needed for the computation of the Lyapunov spectrum. Locally linear approximation was introduced in [45, 46]. We should note that the straight least squares solution of Eq.() is not always optimal and a number of strategies are available to regularize the problem if the matrix becomes nearly singular and to remove the bias due to the errors in the ``independent'' variables. These strategies have in common that any possible improvement is bought with considerable complication of the procedure, requiring subtle parameter adjustments. We refer the reader to Refs. [51, 52] for advanced material.
In Fig. we show iterated predictions of the Poincaré map data from the CO laser (Fig. ) in a delay representation (using nstep in two dimensions). The resulting data do not only have the correct marginal distribution and power spectrum, but also form a perfect skeleton of the original noisy attractor. There are of course artefacts due to noise and the roughness of this approach, but there are good reasons to assume that the line-like substructure reflects fractality of the unperturbed system.
Figure: Time delay representation of 5000 iterations of the local linear predictor nstep in two dimensions, starting from the last delay vector of Fig. .
Casdagli  suggested to use local linear models as a test for nonlinearity: He computed the average forecast error as a function of the neighborhood size on which the fit for and is performed. If the optimum occurs at large neighborhood sizes, the data are (in this embedding space) best described by a linear stochastic process, whereas an optimum at rather small sizes supports the idea of the existence of a nonlinear almost deterministic equation of motion. This protocol is implemented in the routine ll-ar, see Fig. .
Figure: The Casdagli test for nonlinearity: The rms prediction error of local linear models as a function of the neighborhood size . Lower (green) curve: The CO laser data. These data are obviously highly deterministic in m=4 dimensions and with lag =6. Central (blue) curve: The breath rate data shown in Fig. with m=4 and =1. Determinism is weaker (presumably due to a much higher noise level), but still the nonlinear structure is dominant. Upper (red) curve: Numerically generated data of an AR(5) process, a linearly correlated random process (m=5, =1).