CONCLUSIONS

L¹ deconvolutions with IRLS algorithm are especially efficient and robust in the presence of high-amplitude noise bursts. In the general case, their residuals are a sharpened version of the L² residuals, as in predictive deconvolution: the largest amplitudes are increased, the smallest are decreased, and thus the contrast between small and large events is heightened. However, we cannot expect many improvements in the estimation of the phase of the seismic wavelet. Finally, this algorithm could be used for minimizations with mixed L¹-L² norms, and it would thus cope with the simultaneous presence of different kinds of noise.