Showing posts from April, 2014

Testing Matrix Function Algorithms Using Identities

Edvin Deadman and Nick Higham (University of Manchester) write:

In a previous blog post we explained how testing new algorithms is difficult. We discussed the forward error (how far from the actual solution are we?) and the backward error (what problem have we actually solved?) and how we'd like the backward error to be close to the unit roundoff, u.
For matrix functions, we also mentioned the idea of using identities such as sin2A + cos2A = I to test algorithms. In practice, rather than I, we might find that we obtain a matrix R close to I, perhaps with ||R-I|| ≈ 10-13. What does this tell us about how the algorithms for sin A and cos A are performing? In particular, does it tell us anything about the backward errors? We've just written a paper which aims to answer these questions. This work is an output of NAG's Knowledge Transfer Partnership with the University of Manchester, so we thought we'd blog about it here.
Let's consider the identity exp(log A) - A = 0. …