Modified Gauss-Newton full-waveform inversion explained–-why sparsity-promoting updates do matter

Full-waveform inversion can be formulated as a nonlinear least-squares optimization problem. This non-convex problem can be extremely computationally expensive because it requires repeatedly solving large linear systems that correspond to discretized partial differential equations. Randomized subsampling techniques allow us to work with small subsets of (monochromatic) source experiments, reducing the computational cost. However, this subsampling weakens subsurface illumination and introduces subsampling related incoherent artifacts. These subsampling-related artifacts–-in conjunction with local minima that are known to plague full-waveform inversion–-motivate us to come up with a technique to "regularize" this problem. Following earlier work, we take advantage of the fact that curvelets represent subsurface models and perturbations parsimoniously. At first impulse promoting sparsity on the model directly seems the most natural way to proceed, but we will demonstrate that in certain cases it can be advantageous to promote sparsity on the Gauss-Newton updates instead. While constraining the one-norm of the descent directions does not change not change the underlying full-waveform inversion objective, the constrained model updates remain descent directions, remove subsampling-related artifacts and improve the overall inversion result. We empirically observe this phenomenon in situations where the different model updates occur at roughly the same locations in the curvelet domain. We further investigate and analyze this phenomenon, where nonlinear inversions benefit from sparsity-promoting constraints on the updates, by means of a set of carefully selected examples including phase retrieval and full-waveform inversion. In all cases, we observe a faster decay of the residual and model error as a function of the number of iterations.