Open AccessComment on ‘Estimation of resolution and covariance for large matrix inversions’ by J. Zhang and G. A. McMechan
Author(s) -
Deal M. M.,
Nolet G.
Publication year - 1996
Publication title -
geophysical journal international
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.302
H-Index - 168
eISSN - 1365-246X
pISSN - 0956-540X
DOI - 10.1111/j.1365-246x.1996.tb01548.x
Subject(s) - singular value decomposition , mathematics , orthonormal basis , matrix (chemical analysis) , singular value , conjugate gradient method , covariance , covariance matrix , algorithm , resolution (logic) , eigenvalues and eigenvectors , computer science , statistics , physics , materials science , quantum mechanics , artificial intelligence , composite material
When inverting large matrices, iterative techniques are necessary because of their speed and low memory requirements, as opposed to singular value decomposition (SVD). Recently, there have been attempts to obtain information on the quality of the solutions calculated using conjugate gradient (CG) methods such as LSQR. The purpose of this note is to comment on the paper titled “Estimation of resolution and covariance for large matrix inversions’ by Zhang & McMechan (1995), who extend Paige and Saunders’ LSQR algorithm to obtain an orthonormal basis used to approximate resolution and covariance. We show that for larger problems, where the number of orthogonal vectors is several orders of magnitude smaller than the number of model parameters, the vectors obtained do not adequately span the range of the model space. We use a synthetic borehole experiment to illustrate the differences between the singular value spectrum obtained through the more complete method of SVD and the Ritz value spectrum that results from a simple extension of LSQR, We also present a trivial numerical example to illustrate the differences between Zhang & McMechan's approximate resolution matrix and the true resolution.