Corrector-predictor methods for sufficient linear complementarity problems (original) (raw)

We present a new corrector-predictor method for solving sufficient linear complementarity problems for which a sufficiently centered feasible starting point is available. In contrast with its predictor-corrector counterpart proposed by Miao, the method does not depend on the handicap κ of the problem. The method has O((1 + κ) √ nL)-iteration complexity, the same as Miao's method, but our error estimates are sightly better. The algorithm is quadratically convergent for problems having a strictly complementary solution. We also present a family of infeasible higher order corrector-predictor methods that are superlinearly convergent even in the absence of strict complementarity. The algorithms of this class are globally convergent for general positive starting points. They have O((1 + κ) √ nL)-iteration complexity for feasible, or "almost feasible", starting points and O((1 + κ) 2 nL)-iteration complexity for "sufficiently large" infeasible starting points.