1: \begin{abstract} 

2:   \noindent {\rm Following \citet{hartigan}, a cluster is defined as a

3:     connected component of the $t$-level set of the

4:     underlying density, i.e., the set of points for which the density

5:     is greater than $t$. %  The $t$-level set is considered as

6: %     the meaningful part of the support of the underlying distribution.

7:     A clustering algorithm which combines a density estimate with

8:     spectral clustering techniques is proposed.  Our algorithm is

9:     composed of two steps.  First, a nonparametric density estimate is

10:     used to extract the data points for which the estimated density

11:     takes a value greater than $t$.  Next, the extracted

12:     points are clustered based on the eigenvectors of a graph

13:     Laplacian matrix.  Under mild assumptions, we prove the almost

14:     sure convergence in operator norm of the empirical graph Laplacian

15:     operator associated with the algorithm.  Furthermore, we give the

16:     typical behavior of the representation of the dataset into the

17:     feature space, which establishes the

18:     strong consistency of our proposed algorithm.

19:     \\

20: 

21:     \noindent \emph{Index Terms}: Spectral clustering, graph, unsupervised

22:     classification, level sets, connected components.

23:     % \\

24: 

25: %     \noindent \emph{AMS 2000 Classification}: 62G08, 62G05.  

26:   }

27: 

28: \end{abstract}

29: