1: \begin{abstract}

2: Spectral kernel methods are techniques

3: for transforming data into a coordinate system that

4: efficiently reveals the geometric structure---

5: in particular, the ``connectivity''---of the data.

6: These methods depend on certain tuning parameters.

7: We analyze the dependence of the method on these tuning parameters.

8: We focus on one particular technique---diffusion maps---but our analysis

9: can be used for other methods as well.

10: We identify the population quantities implicitly being estimated,

11: we explain how these methods relate to classical kernel smoothing and

12: we define an appropriate risk function for analyzing the estimators.

13: We also show that, in some cases, fast rates of convergence

14: are possible even in high dimensions.

15: \end{abstract}

16: