An investigation of the solution of least squares problems using the QR factorisation
Supervisor
Suitable for
Abstract
Experimental data inevitably contains error. These experimental observations are often compared to theoretical predictions by writing as a least squares problem, i.e. minimising the sum of squares between the experimental data and theoretical predictions. These least squares problems are often solved using a QR-factorisation of a known matrix, which uses the Gram-Schmidt method to write the columns of this matrix as a linear sum of orthonormal vectors. This method, when used in practice, can exhibit numerical instabilities, where the (inevitable) numerical errors due to fixed precision calculations on a computer are magnified, and may swamp the calculation. Instead, a modified Gram-Schmidt method is used for the QR-factorisation. This modified Gram-Schmidt factorisation avoids numerical instabilities, but is less computationally efficient. The first aim of this project is to investigate the relative computational efficiencies of the two methods for QR-factorisation. The second aim is to use the QR-factorisation to identify: (i) what parameters can be recovered from experimental data; and (ii) whether the data can automatically be classified as "good" or "bad".