Convergence of quasi-Newton matrices generated by the symmetric rank one update (Q1176575): Difference between revisions

Quasi-Newton algorithms for unconstrained nonlinear minimization are update methods. In every update of those algorithms an approximation of the objective function's Hessian matrix is generated to constitute the line search formula. It is still an open problem if for the whole class of quasi-Newton algorithms under reasonable conditions the sequence of generated matrices converges and, furthermore, converges to the true Hessian matrix. This paper shows that for the symmetric rank one update formula the generated matrices converge globally to the true Hessian matrix. The rate of convergence is also considered. The authors give numerical experiments to confirm their theoretical results. Those numerical experiment also show the convergence of the Hessian approximations to be substantially slower for other known quasi-Newton methods.