Figure 4:

Comparison of infinitesimal and finite lag time loss functions. (left) RMS error in the committor for the Müller-Brown dynamics mapped to the Swiss roll obtained with the infinitesimal lag time loss function in (22) as the frequency of the sinusoidal perturbation, $\omega$, is increased. The other panels show the error as the lag time is increased with the frequency fixed at $\omega =0$ (center) or $\omega =10$ (right). Shading shows the standard deviation in the error from training on ten independent data sets.