Figure 2:

Comparison with Galerkin methods. (left) For an MSM estimate with $k=300$ clusters, the root mean square (RMS) error in the committor for the Müller-Brown model as the Swiss roll is wound tighter (higher $f$ in (17)). Shading shows the standard deviation in the error from training on ten independent data sets. (right) Number of MSM clusters needed to achieve an RMS error in the committor of less than 0.045.