A convergent dynamic method for large minimization problems (Q1124283)

A new gradient algorithm for unconstrained minimization: min F(x), \(x\in R^ n\), requiring no line searches and consequently no function evaluations, is developed. This algorithm approaches the original problem by the associated dynamic problem: solve the equations of motion \(x''(t)=-F(x(t))\) subject to initial conditions \(x(0)=x_ 0\), \(x'(0)=v_ 0=0\) \((x_ 0\) starting point). An approximate solution of this equations is numerically obtained by applying the so-called ``leap-frog'' method. For convex functions the convergence of the given method is proved. The performance of this algorithm is compared with the conjugate gradient algorithm of \textit{M. J. D. Powell} [Math. Programming 12, 241-254 (1977; Zbl 0396.90072)].