Same as Fig. 1 for a mixture of two systematic effects sharing some partial correlation. The individual LBL RVs are shown in the top row, coloured according to wavelength (left) or y-pixel position (right). The two systematic effects V₁ and V₂ (shown by the red arrows in the top left panel) depend, respectively on these two variables. A classical PCA decomposition results in orthogonal components PC₁ and PC₂ (as shown by the arrows in the top left panel), that are unrelated to the main systematics effects. In the bottom row, we first group-average the LBL RVs according to wavelength (left) and y-position (right) before performing PCA on the result to identify principal components PC₁ and PC′₁that are proportional to V₁ and V₂, respectively. The same would occur for a cloud of data with outliers.