Matrix-free quadratic-penalty methods for PDE-constrained optimization

The large scale of seismic waveform inversion makes matrix free implementations essential. We show how to exploit the quadratic penalty structure to construct matrix free reduced-space and full-space algorithms, which have some advantages over the commonly used Lagrangian based methods for PDE-constrained optimization. This includes the construction of effective and sparse Hessian approximations and reduced sensitivity to the initial guess. A computational bottleneck is the need to solve a large least squares problem with a PDE block. When direct solvers are not available, we propose a fast matrix free iterative approach with reasonable memory requirements. It takes advantage of the structure of the least squares problem with a combination of preconditioning, low rank decomposition and deflation.