Rounding error analysis of the classical Gram-Schmidt orthogonalization process (Q2486676)

Given an \(m\times n\) matrix \(A\) of full column rank, this paper is concerned with the numerical properties of the classical Gram-Schmidt process (CGS) for computing an orthonormal basis of the space spanned by the columns of \(A\). Based on a backward error bound for the upper triangular factor produced by CGS, it is proven that the loss of orthogonality of the computed basis can be bounded roughly by the machine precision multiplied with the squared condition number of \(A\). Applying CGS once again to the computed basis is shown to make the squared condition number disappear, provided that \(A\) is also numerically of full rank. This verifies the convential wisdom that applying CGS twice is enough to obtain a numerically orthogonal basis.