It has long been known that the method of time-delay embedding can be used to reconstruct nonlinear dynamics from time series data. A less-appreciated fact is that the induced geometry of time-delay coordinates increasingly biases the reconstruction toward the stable directions as delays are added. This bias can be exploited, using the diffusion maps approach to dimension reduction, to extract dynamics on desired time scales from high-dimensional observed data. We demonstrate the technique on a wide range of examples, including data generated by a model of meandering spiral waves and video recordings of a liquid-crystal experiment.1. Introduction. The method of time-delay embedding was first introduced by Takens, Ruelle, and others for the purpose of reconstructing dynamical attractors from data [32,23,1,14]. With a series of generic delayed observations, it was shown that topological properties are preserved in reconstruction dimensions greater than 2d, for a smooth attractor of dimension d [32], and for an attractor of fractal dimension d [26]. Considerable effort followed to develop methods of estimating attractor dimension, in order to carry out embeddings in a Euclidean space of minimal dimension.More recently, the development of pervasive and cheap sensors has caused a shift in emphasis toward methods capable of handling large quantities of time-ordered data. For example, we could imagine a complex system with a relatively low-dimensional attractor, possibly chaotic, where the observations are represented by a high-dimensional multivariate time series. In such a case, it is of great interest to apply a data analysis technique to a video of an experiment, for instance, and to seek ways to selectively project that data onto various dynamical time scales of interest.Snapshots of spatiotemporal patterns produced by electroconvection in a liquid crystal are shown in Figure 1. Simulations of complex spatiotemporal models, such as Rayleigh-Bénard convection, indicate that there may be a low-dimensional representation of the process [15]. However, conventional techniques of dimensionality reduction such as the Karhunen-Loeve decomposition have been unable to recover a low-dimensional process even for low driving