Evaluation of data compression techniques for the inference of stellar atmospheric parameters from high-resolution spectra

The determination of stellar atmospheric parameters from spectra suffers the so-called curseof- dimensionality problem, which is related to the higher number of input variables (flux values) compared to the number of spectra available to fit a regression model (this collection of examples is known as the training set). This work evaluates the utility of several techniques for alleviating this problem in regression tasks where the objective is to estimate the effective temperature (T-eff), the surface gravity (log g), the metallicity ([M/H]) and/or the alpha-to-iron ratio ([alpha/Fe]). The goal of the techniques analysed here is to achieve data compression by representing the spectra with a number of variables much lower than the initially available set of fluxes. The experiments were performed with high-resolution spectra of stars in the 4000-8000 K range for different signal-to-noise ratio (SNR) regimes. We conclude that independent component analysis (ICA) performs better than the rest of techniques evaluated for all SNR regimes. We also assess the necessity to adapt the SNR of the spectra used to fit a regression model (training set) to the SNR of the spectra for which the atmospheric parameters are needed (evaluation set). Within the conditions of our experiments, we conclude that at most only two such regression models are needed (in the case of regression models for effective temperatures, those corresponding to SNR = 50 and 10) to cover the entire SNR range. Finally, we also compare the prediction accuracy of effective temperature regression models for increasing values of the training grid density and the same compression techniques.