Purdue University

EAS 657

Geophysical Inverse Theory

Robert L. Nowack

Lecture 8

Factorizations of Ax = y

 

            For a set of equations Ax = y, choose a nonsingular matrix M such that

 

MAx = My

 

where MA is triangular.  We want to find MA = R where

 

 

where U is square upper triangular matrix.  Since M is nonsingular then A = M-1R.

 

            Ex)  A square matrix can be factored into a lower and upper triangular matrix

 

A = LU

 

This is for rapid computation of a square non-singular system Ax = y using Gaussian elimination.

 

            Ex)  For purely overdetermined systems, then we can solve this by constructing the “normal equations”

 

A*Ax = A*y

 

Where A*A is symmetric and positive definite (assuming the eigenvalues ).  Then,  can be factored as

 

 

where L is a lower triangular matrix and has one’s along the main diagonal, and D is a diagonal matrix with the eigenvalues of .  This is called the Cholesky factorization of a square symmetric matrix.  This can then be used to solve by backsubstitution the normal equations resulting from least squares.

 

 

            Ex)  The QR factorization

 

This is the preferred way to solve the purely overdetermined system.  This will result in the least squares solution and works directly with the original matrix solution instead of the normal equations which may have numerical stability problems due to squared singular values.

 

            The idea is to decompose A = Q*R, where Q is a matrix product of a series of orthogonal transformations and is unitary and

 

 

where U is a square upper triangular matrix.  Now,

 

 

where QA is (MxM) (MxM) and R is (MxN).  Each H is called a householder transformation.  It is an orthogonal matrix of the form

 

 

Since  and

 

 

Thus H is an orthogonal matrix.  When we use a householder matrix to transform a vector a to another vector b, then

 

 

 

Thus,

 

 

            Ex)  Find H such that .  Let

 

 

where

 

 

Then,

 

 

 

with

 

 

Then,

 

 

This is the first step of the QR alrogithm where we let

 

 

then,

 

 

where

 

 

            Now the transformations Hi are designed to zero out the elements i + 1 to M of the ith column of a partially triangularized matrix without altering the first (i-1) columns.  Thus,

 

 

Thus,

 

 

and

 

 

            The above algorithm assumes that the A matrix is not rank deficient with P=N.  If P < N, then the algorithm must be modified somewhat, with

 

 

where Q is MxM, A is MxN and V is a series of householder transformation applied on the right and is an NxN orthogonal matrix.  In fact, QR type algorithms are usually first steps in SVD algorithms.  For SVD, the final numerical step is to reduce RPxP to .  But this final step is comparatively expensive and in many cases may not be needed.

 

            The most common application of the QR algorithm is to the least squares or damped least squares problem assuming a full column rank.  In the least squares problem, one minimizes

 

 

 

Note:  length is preserved by orthogonal matrices.  Then,

 

 

Let

 

 

The error will be minimized by solving by back substitution for x

 

 

            Ex)  Damped least squares

 

 

For this case, do a Q-R decomposition on the matrix  for

 

 

            The primary advantage of solving least squares problems with QR algorithms is not speed, but the numerical stability in dealing with a system with eigenvalues  instead of the normal equations with eigenvalues .

 

 

 

Operation Count for Solving Least Squares Problems with M  N

 

 

 

Algorithm                                               Flop count 1, add + 1 multiply

 

Normal Equations                                                  

 

Q-R Householder Orthogonalization                      

 

SVD