The bit-cost of some algorithms for the solution of linear systems

We present an analysis of the bit-cost of some numerical linear system solvers. We use measures of the computational cost of algorithms, which are deeply related to their numerical behaviour. We derive upper bounds to the worst case bit-performance of the Gaussian elimination, Jacobi's and Newton's methods, implemented either in a sequential or in a parallel environment. Moreover, we analyze an interesting special case, e.g. the solution of triangular Toeplitz linear systems. © 1988.