Abstract
Examples are given in which quasi-Newton and other conjugate descent algorithms fail to converge to a minimum of the object function. In the first there is convergence to a point where the gradient is infinite; in the second, a region characterized by a fine terraced structure causes the iterates to spiral indefinitely. A modification of the second construction gives convergence to a point where the gradient is non-zero, but the gradient is not continuous at this point. The implication of these examples is discussed.
Similar content being viewed by others
Reference
M.J.D. Powell, “Some properties of the variable Metric Algorithm”, in: F.A. Lootsma, ed.,Numerical Methods for non-linear optimization (Academic Press, New York 1972) pp. 1–17.
Author information
Authors and Affiliations
Rights and permissions
About this article
Cite this article
Thompson, J.R. Examples of non-convergence of conjugate descent algorithms with exact line-searches. Mathematical Programming 12, 356–360 (1977). https://doi.org/10.1007/BF01593803
Received:
Revised:
Issue Date:
DOI: https://doi.org/10.1007/BF01593803