Sparse quasi-Newton methods and the continuation problem
The problem of tracing a smooth path arises in many engineering problems, the solution of parametric differential equations and eigenvalue problems; it also finds application in the solution of nonlinear systems of equations by homotopy techniques. In many instances, the path is defined implicitly as the solution of a system of equations whose Jacobian matrix is large and sparse. Robust simplicial path-following techniques cannot be applied to large problems since the work involved rises rapidly with increasing dimension. This dissertation addresses the numerical problems involved in tracing the path for large sparse systems by the use of a predictor-corrector algorithm. The corrector phase of a predictor-corrector algorithm is very expensive if Newton's method is used as the corrector. We investigate the use of sparse quasi-Newton techniques to reduce this expense. In order to avoid the drawbacks of the sparse Broyden method - the need for a matrix factorization on each iterate and the need to store the Jacobian matrix and its factors - we examine techniques for directly updating the LU factors of the approximation to the Jacobian matrix. Under reasonable assumptions on the systems of equations to be solved, a proof of local Q-superlinear convergence is presented for two sparse updating techniques. A predictor-corrector algorithm employing these sparse updating techniques is implemented in a Fortran code and numerical results are obtained demonstrating the advantages to be gained from the use of quasi-Newton methods for the large sparse continuation problem.
- Research Organization:
- Stanford Univ., CA (USA). Systems Optimization Lab.
- DOE Contract Number:
- AS03-76ER72018
- OSTI ID:
- 5403490
- Report Number(s):
- SOL-85-8; ON: DE85015390
- Country of Publication:
- United States
- Language:
- English
Similar Records
Performance analysis of nonlinear equation solution techniques for chemical process design
A STRUCTURED QUASI-NEWTON ALGORITHM FOR OPTIMIZING WITH INCOMPLETE HESSIAN INFORMATION