We propose a spavsity preserving algorithm for solving large-scale, nonlineav programming problems. The algorithm solves at each iteration a subproblem, which contalns a lineavized objective function augmented by a simple quadratic term and linearized constraints. The quadratic term added to the lineavized objective function plays the role of step restriction which is essential in ensuring global convergence of the algorithm. Ir the conjugate gradient method or successive over-relaxation method is used to solve the subproblems, the spavsity of the original problem is preserved, because those methods only require simple operations on the rows of the constraint matrix. Thus, lavge-scale problems can be dealt with when the constraint matrices ave sparse enough to be stored in a compact form. Praztical implementation of the algorithm is described and computational results ave reported.Large-scale nonlinear programming problems usually have the property that many of their problem functions are linear and sparse. Thus, successive linear programming (SLP) algorithms [4,12] have been preferred for solving them fora long time, in spite of the lack of theoretical convergence properties. In a recent paper [17], Zhang, Kim and Lasdon propose a globally convergent SLP algorithm in which an exact penalty function technique is incorporated with a trust region method. Though SLP algorithms can solve very large problems [1], they may not completely preserve the sparsity which the original problems may have, because of possible fill-in caused in solving LP subproblems.The purpose of this paper is to propose a sparsity preserving algorithm, which is also globally convergent to an optimal solution of the problem. The proposed algorithm solves at each iteration a subproblem which contains a linearized objective function augmented by a simple quadratic term and linearized constraints. Though this algorithm resembles successive quadratic programming (SQP) algorithms, it is based on an idea entirely different from typical SQP methods such as [3,5,13,14]. In our algorithm, the quadratic term added to the linearized objective