Phase transition of graph Laplacian of high dimensional noisy random point cloud
share
We systematically explore the spectral distribution of kernel-based graph Laplacian constructed from high dimensional and noisy random point cloud in the nonnull setup. An interesting phase transition phenomenon is reported, which is characterized by the signal-to-noise ratio (SNR). We quantify how the signal and noise interact over different SNR regimes; for example, how signal information pops out the Marchenko-Pastur bulk. Motivated by the analysis, an adaptive bandwidth selection algorithm is provided and proved, which coincides with the common practice in real data. Simulated data is provided to support the theoretical findings. Our results paves the way towards a foundation for statistical inference of various kernel-based unsupervised learning algorithms, like eigenmap, diffusion map and their variations, for real data analysis.
READ FULL TEXT