Inexact Hessian-vector products in reduced-space differential-equation constrained optimization

Reduced-space inexact-Newton–Krylov (RSNK) algorithms provide a modular and scalable framework for solving differential-equation-constrained optimization problems; thus, this class of algorithms provides an attractive compromise between poorly scaling “black-box” methods and more intrusive, full-space optimization algorithms. One of the challenges of implementing RSNK methods is the efficient solution of the linear subproblems, which involve the reduced Hessian. This paper explores inexact Hessian-vector products to improve the efficiency of these subproblems. The reduced-Hessian-vector products, which are required by the Krylov solver in RSNK, can be determined by solving two second-order adjoint equations. To reduce computational cost, we consider the approximate (i.e. inexact) solution of the second-order adjoints. We present bounds for the second-order adjoint tolerances that ensure the Hessian-vector product remains sufficiently accurate for inexact-Krylov methods. The bounds involve the 2-norm of the state and first-order adjoint sensitivities, and we provide an inexpensive Lanczos algorithm to estimate these matrix norms. Using numerical experiments, we investigate the proposed RSNK algorithm and compare it to other optimization algorithms.