Calibration of α₁ and α₂ as a function of the spectrum S/N for the erf Fourier filter. Left: high S/N (∼12 000) spectrum of α Cen B (blue curve) is considered to be noise free. Different realisations of Poisson noise are generated (grey curves) in order to determine the best pair (α₁, α₂) whose corresponding smoothed curve (red curve) is most similar to the original curve. Only a subpart of the spectra is displayed for graphical considerations. Middle: (α₁, α₂) parameter space for different S/N realisations. The simulations were run on a grid of evenly spaced values of α₁ and α₂. The colours encode the metric used here, i.e. the standard deviation between the smoothed and noise-free spectrum. The local minimum of the parameter space is indicated by a yellow cross, and the contour level of the 5% best simulations is drawn in white. The noisier the spectrum, the lower the best α₁. The best value for α₂ appears rather constant around 0.175. Right: fourier transform of the spectra of the middle column (blue curve). The high-frequency noise (grey curve) increases significantly from S/N = 740 to S/N = 25 such that the centre of the erf function moves towards shorter frequencies. The erf filter corresponding to the best (α₁, α₂) pair is displayed (black curve). The red curve is the Fourier transform after smoothing by the function. The typical frequency scale σ⁻¹, where the sigma is the width of the Gaussian fitted to the stellar CCF, is shown as a dashed vertical line.