1: \begin{abstract}

2: 

3: 

4: \noindent

5: Kernelized Gram matrix $W$ constructed from data points $\{x_i\}_{i=1}^N$ as $W_{ij}= k_0(  \frac{ \| x_i - x_j \|^2} {\sigma^2} ) $

6: is widely used in graph-based geometric data analysis

7: and unsupervised learning. 

8: An important question is how to choose the kernel bandwidth $\sigma$,

9: and a common practice called self-tuned kernel adaptively sets a $\sigma_i$ at each point $x_i$ by the $k$-nearest neighbor (kNN) distance.

10: When $x_i$'s are sampled from a $d$-dimensional manifold embedded in a possibly high-dimensional space,

11: unlike with fixed-bandwidth kernels,

12: theoretical results of graph Laplacian convergence with self-tuned kernels

13: %\old{,however,}

14: have been incomplete. 

15: This paper proves the convergence of graph Laplacian operator $L_N$

16: to manifold (weighted-)Laplacian for

17: a new family of kNN self-tuned kernels $W^{(\alpha)}_{ij}

18:  = k_0( \frac{  \| x_i - x_j \|^2}{ \epsilon \hat{\rho}(x_i) \hat{\rho}(x_j)})/\hat{\rho}(x_i)^\alpha \hat{\rho}(x_j)^\alpha$, 

19: where $\hat{\rho}$ is the estimated bandwidth function {by kNN},

20: and the limiting operator is also parametrized by $\alpha$.

21: When $\alpha = 1$, the limiting operator is the weighted manifold Laplacian $\Delta_p$.

22: Specifically, we prove the  point-wise  convergence of $L_N f $ and convergence of the graph Dirichlet form with rates.

23: Our analysis is based on first establishing 

24: a $C^0$ consistency for 

25: $\hat{\rho}$

26: which bounds the relative estimation error  $|\hat{\rho} - \bar{\rho}|/\bar{\rho}$ uniformly 

27: with high probability, 

28: where $\bar{\rho} = p^{-1/d}$, and $p$ is the data density function. 

29: Our theoretical results reveal the advantage of self-tuned kernel over fixed-bandwidth kernel 

30: via smaller variance error in low-density regions.

31: In the algorithm, no prior knowledge of $d$ or data density is needed. 

32: The theoretical results are supported by numerical experiments  on simulated  data and hand-written digit image data.

33: \end{abstract}

34: