On the complexity of steepest descent, Newton's and regularized Newton's methods for nonconvex unconstrained optimization problems

摘要

It is shown that the steepest-descent and Newton's methods for unconstrained nonconvex optimization under standard assumptions may both require a number of iterations and function evaluations arbitrarily close to O(ε ~(?2)) to drive the norm of the gradient below ε. This shows that the upper bound of O(ε ~(?2)) evaluations known for the steepest descent is tight and that Newton's method may be as slow as the steepest-descent method in the worst case. The improved evaluation complexity bound of O(ε ~(?3/2)) evaluations known for cubically regularized Newton's methods is also shown to be tight.