Backward Stability of Householder and Givens QR

Backward Stability of Householder and Givens QR#

This is perhaps the most important theoretical guarantee for these methods. While rounding errors mean we don’t compute the exact \(Q\) and \(R\) for a given \(A\), we can prove something stronger.

Theorem 18 (Backward Stability of QR Factorization)

When QR factorization is computed using a sequence of Householder or Givens transformations in floating-point arithmetic, the computed factors \(\hat{Q}\) and \(\hat{R}\) are the exact orthogonal and upper-triangular factors for a slightly perturbed matrix \(A + \delta A\). The size of the perturbation is small, bounded by:

\[ \frac{\|\delta A\|}{\|A\|} = O(\epsilon_{\text{machine}}) \]

where \(\epsilon_{\text{machine}}\) is the machine precision.

Why it’s important: This result provides a rock-solid guarantee. It tells us that the algorithm gives the right answer for a problem that is very close to the one we started with. For most real-world problems where the input data \(A\) has some inherent uncertainty, this is a fantastic result. It’s the gold standard for numerical algorithms.