The impact of numerical derivatives on radial velocity extraction

Abstract

The radial velocity (RV) method is a widely used technique to infer planetary masses and orbital parameters. One of the most widely used methods to compute RVs is based on the alignment of a high-SNR, data driven, stellar model with individual observations, commonly referred to as template matching. Typically, the alignment is performed through a $\chi^2$ minimization, but some approaches rely on the derivative of the stellar template to do so, which is often the case in line-by-line methods. In this paper we aim to explore the limitation of derivative-based methods for RV extraction in the case of using low-SNR stellar models. We use simulated Gaussian profiles to investigate the effect of computing a numerical derivative of the stellar template, in comparison with the usage of an analytical profile. The impact on RV and associated uncertainty is then analyzed as a function of the signal-to-noise ratio (SNR) of the line. Then, using real observations we compare the residuals between a derivative-based RV extraction and a classical template matching implementation, as function of the SNR of the stellar template. We find that on simulated Gaussian profiles the usage of the numerical approach leads to a RV residual at the level of the meter per second, at a per-pixel SNR regime of 100. An increase of the SNR leads to a decrease of this impact, falling under the current noise-floor of state-of-the-art spectrographs at SNR>1000 for a single spectral line. The inclusion of multiple spectral lines in the simulations lead to an overall decrease of the contamination, across all SNR regimes. The application to a real dataset presents a decrease of the RV impact with the increase of the template's SNR, albeit still presenting a 12.5 cm/s difference when including 78 observations in the stellar template.