1: \begin{abstract}

2: Solving large-scale systems of nonlinear equations/inequalities is a

3: fundamental problem in computing and optimization. In this paper, we propose

4: a generic successive projection (SP) framework for this problem.

5: The SP sequentially projects the current iterate onto the constraint set

6: corresponding to each nonlinear (in)equality. It extends

7: von Neumann's alternating projection for finding a point in the

8: intersection of two linear subspaces, Bregman's method for

9: finding a common point of convex sets and the Kaczmarz method

10: for solving systems of linear equations to the more general case of multiple nonlinear

11: and nonconvex sets. The existing convergence analyses on randomized Kaczmarz

12: are merely applicable to linear case. There are no theoretical convergence results of the SP

13: for solving nonlinear equations. This paper presents the

14: first proof that the SP locally converges to a solution of nonlinear

15: equations/inequalities at a linear rate. Our work establishes

16: the convergence theory of the SP for the case of multiple nonlinear and nonconvex sets.

17: Besides cyclic and randomized projections, we devise two new

18: greedy projection approaches that significantly

19: accelerate the convergence. Furthermore, the theoretical

20: bounds of the convergence rates are derived. We reveal that the convergence

21: rates are related to the Hoffman constants of the Jacobian matrix

22: of the nonlinear functions at the solution. Applying the SP to solve the graph realization problem,

23: which attracts much attention in theoretical computer science, is discussed.

24: \end{abstract}

25: