Linear Point Standard Ruler Estimation with Neural Networks

The linear point of the two-point correlation function, the midpoint between the peak and dip of the BAO feature, can be used as a standard cosmological ruler and has been shown to be more stable to non-linear evolution than the BAO peak itself. The scales of the dip and peak are typically measured by fitting polynomials to the correlation function and calculating the critical points; however, in certain cases, such as cosmologies far from the fiducial or surveys and simulations with a limited volume, the BAO feature may be missing or lost to noise, so a polynomial estimator would fail to detect a linear point scale. We present a new scheme to estimate the linear point using neural networks. Trained on correlation functions from the Quijote simulations spanning a wide range cosmological parameters with different linear point scales, the network performs as well as the polynomial estimator, but also recovers similarly accurate linear points from the significant fraction of correlation functions for which the polynomial estimator fails to detect the linear point. We also test the robustness of the model against constant additive terms and redshift space distortions.