Dimensionality Reduction in Nonlinear Optical Datasets via Diffusion Mapping: Case Study of Short-Pulse Second Harmonic Generation

摘要

We have studied the application of the diffusion mapping technique to dimensionality reduction and clustering in multidimensional optical datasets. The combinational (input-output) data were obtained by sampling search spaces related to optimization of a nonlinear physical process, short-pulse second harmonic generation. The diffusion mapping technique hierarchically reduces the dimensionality of the data set and unifies the statistics of input (the pulse shape) and output (the integral output intensity) parameters. The information content of the emerging clustered pattern can be optimized by modifying the parameters of the mapping procedure. The low-dimensional pattern captures essential features of the nonlinear process, based on a finite sampling set. In particular, the apparently parabolic two-dimensional projection of this pattern exhibits regular evolution with the increase of higher-intensity data in the sampling set. The basic shape of the pattern and the evolution are relatively insensitive to the size of the sampling set, as well as to the details of the mapping procedure. Moreover, the experimental data sets and the sets produced numerically on the basis of a theoretical model are mapped into patterns of remarkable similarity (as quantified by the similarity of the related quadratic-form coefficients). The diffusion mapping method is robust and capable of predicting higher-intensity points from a set of low-intensity points. With these attractive features, diffusion mapping stands poised to become a helpful statistical tool for preprocessing analysis of vast and multidimensional combinational optical datasets.