A quasi-Gauss-Newton method for solving nonlinear algebraic equations (Q1203697)

The classical Gauss-Newton method for solving a system of nonlinear equations \(F(x)= 0\) requires the solution of the normal equations \(J^ T Jp= -J^ T F\) in each step, where \(J\) denotes the Jacobian. The authors propose to replace the Jacobian \(J\) by a rank-one least change secant approximation \(B\), e.g. the Broyden update. Then the next approximation \(B^ T_{k+1} B_{k+1}\) for \(J^ T J\) is a symmetric rank-two modification of the previous approximation \(B^ T_ k B_ k\). An efficient Cholesky factorization update technique, well-known for factorizing Hessians in unconstrained minimization problems, is applied to \(B^ T_{k+1} B_{k+1}\). The resulting algorithm is significantly less costly compared with a classical rank-one quasi-Newton method with sequential QR factorization. The question of ill-conditioning by using a method based on the normal equations is discussed. Computational results illustrate the superiority of the proposed method for a number of test problems.