Variance parameters estimation - application to full waveform inversion

Many inverse problems include nuisance parameters. While not of direct interest, these parameters are required to recover primary parameters. In order to estimate these nuisance parameters as well as the primary parameters in large-scale inverse problems, a method based on variable projection, which consists in projecting out a subset over the variables, has been developed. We present here the application of this method to the problem of variance parameters estimation in multiple datasets, which is an important problem in many areas including geophysics. More precisely, we apply the method to Full Waveform Inversion and demonstrate the improvement in recovery of the model parameters in the case where the variance of the noise increases with the frequency.