Nesterov Majorization

Suppose is a function with third derivatives that are bounded in the sense that It is sufficient for this that the Hessian is Lipschitz continuous with Lipschitz constant . Majorization based on was first discussed in an important article by Nesterov and Polyak [2006]

Under these conditions we have the majorizer The term is convex in , but the majorizer itself is generally not convex, although it is convex whenever is. Majorizer has continuous first derivatives and for the second derivative is It follows that and thus, by the squeeze theorem,


The majorization algorithm is, as usual, and the majorizer is minimized at a point where . We write the stationary equation as We write this as two equations, using what is effectively a form of decomposition. Let be an eigen decomposition, and define and . Then solving for and using gives the secular equation in


The Cartesian folium is Thus with If then and thus This gives the majorization we are looking for The derivatives are with . Thus or Thus we must have


Thus majorizes at . Now and where .


Suppose is a multivariate quartic with bounds on the third and fourth derivatives. Thus We minimize over , as usual. Thus we must solve Expand this to the two equations