An estimate of the speed of convergence for certain methods of coordinate descent (Q1086172)

We analyze the convergence speed of certain coordinate descent methods for the solution of an unconditional minimization problem in the form \(\phi\) (x)\(\to \min\), \(x\in E_ n\), where \(E_ n\) is a Euclidean space of dimension n and \(\phi\) (x) is a convex function lying in the class \(C^{1,1}(E_ n)\), i.e., the gradient \(\phi\) '(x) satisfies the Lipschitz condition \(\| \phi '(x)-\phi '(y)\| \leq L\| x-y\|\). Along with linear methods of coordinate descent of the form (1) \(x_{k+1}=x_ k-\beta_ kS_ k\), \(k=0,1,...\), where \(S_ k\) is some coordinate direction of descent and \(\beta_ k\) is the step size in this direction, nonlinear methods of coordinate descent are considered with a general iteration scheme given by the equation \(x_{k+1}=x_ k-a_ kS_ k-\beta_ kd_ k\), where \(S_ k\) and \(d_ k\) are certain coordinate descent directions, \(\alpha_ k\), \(\beta_ k\) are the components of the step in these directions.