A GCV based method for nonlinear ill-posed problems (Q1977979)

The authors deal with the solution of the nonlinear ill-posed equation \(F(m)=b\), where \(F:\mathbb{R}^M \to\mathbb{R}^N\) is a differentiable operator with \(M >N\), and right-hand side \(b\) is noisy. Let \(\phi(\beta, m)=\|F(m)-b\|^2+ \beta \|W(m-m_{\text{ref}}) \|^2\) be the Tikhonov functional of the problem. The standard Gauss-Newton iterative procedure \(m_{k+1}=m_k + \delta m_k\), \[ (F^{\prime T}(m_k) F''(m_k)+ \beta W^T W) \delta m_k= F^{\prime T}(m_k)(b-F(m_k))- \beta W^T W (m_k-m_{\text{ref}}) \] is modified by a stepsize strategy \(m_{k+1}= m_k +\omega_k \delta m_k\), \(\omega_k \in (0,1)\) such that \(\phi(\beta, m_{k+1}) <\phi(\beta, m_k)\). The regularization parameter \(\beta\) is suggested to choose by generalized cross validation (CGV) technique previously developed for linear equations. The obtained heuristic method for the original equation is tested on discrete magnetotelluric and gravimetric problems.