This letter shows that in a repeat-pass data set of synthetic aperture radar (SAR) images, a long-term coherent component, when present, can be recovered by coherent temporal filtering of the SAR images and can successively form interferograms with higher signal-to-noise ratio. The validity of the idea is confirmed through simulations and one example with real TerraSAR-X data. The theoretical necessity of using long-term interferograms is also discussed and linked to autoregressive processes, starting from the observation that the optimal weighting is given by the inverse of the covariance matrix.