The impact of interpolation in high-resolution spectroscopy -- The overlooked role of interpolation in radial velocity extraction

Abstract

We explore the impact of spectral interpolation in radial velocity (RV) time-series extracted through template-based methods. We build synthetic datasets with Gaussian profiles to evaluate flux residuals and line asymmetry that are a result from changing the sampling location of the lines. We generate synthetic spectra as a sum of Gaussian functions whose parameters were determined through an observed spectrum. The s-BART pipeline was applied to them, allowing to evaluate any biases in RV extraction introduced by its internal assumptions in line shape. Lastly, we apply the s-BART pipeline to ESPRESSO observations of four stars: two that use high-cadence observations over a single night, and two that have observations spread over multiple nights. When extracting RVs from stellar spectra, we change the interpolation algorithm, used in the process of constructing the stellar template and, afterwards, during RV extraction, comparing them with RVs extracted with a widely-used cubic-spline interpolation. We find that synthetic datasets reveal systematic biases with the largest peak-to-peak amplitudes reaching $\sim$ 20 m/s in low SNR cases, with the amplitude decreasing as the SNR of the spectra increases. In the extreme case of noise-free data, we still recover a systematic bias, albeit at the mm/s level, significantly smaller than the RV precision of state-of-the-art instruments. With real observations we find that those from high-cadence observations with small BERV variation are impacted by the choice of the interpolation algorithm. This impact is smaller in higher-SNR cases, where the peak-to-peak amplitude reaches $\sim$ 1 m/s. In the comparatively lower-SNR case we find peak-to-peak residuals as large as $\sim$ 25 m/s . In cases where the observations are spread over a larger BERV window, we find an upper limit of 20 cm/s of RV scatter for this systematic signal.