1: \begin{abstract}

2: 		Given 

3: 		%a regular version of 

4: 		the joint distribution of two random variables $X,Y$ 

5: 		on

6: 		some second countable locally compact Hausdorff space,

7: 		we investigate the statistical approximation of 

8: 		the $L^2$-operator $\ko$ defined by 

9: 		${[ \ko f](x) := \Ex[ f(Y) \mid X = x ]}$

10: 		under minimal assumptions.

11: 		By modifying its domain, we

12: 		prove that $\ko$ can be arbitrarily well approximated in operator norm

13: 		by Hilbert--Schmidt operators acting on a

14: 		reproducing kernel Hilbert space.

15: 		This fact allows to

16: 		estimate $\ko$ uniformly 

17: 		by finite-rank operators over a dense subspace

18: 		even when $\ko$ is not compact. 

19: 		In terms of modes of convergence,

20: 		we thereby obtain the superiority of

21: 		kernel-based techniques over classically used

22: 		parametric projection approaches such as Galerkin methods. 

23: 		This also provides a novel perspective on which

24: 		limiting object the nonparametric estimate of $\ko$ converges to.

25: 		As an application, 

26: 		we show that these results are particularly important

27: 		for a large family of spectral analysis techniques for Markov

28: 		transition operators.

29: 		Our investigation also gives a 

30: 		new asymptotic perspective on the so-called

31: 		kernel conditional mean embedding, which is the theoretical foundation

32: 		of a wide variety of techniques in kernel-based nonparametric inference.	

33:    \end{abstract}

34: