For any square matrix A, consider the equation det(A-lambda*I)=0. This is a polynomial equation of degree N in the variable lambda. Hence there are exactly N complex roots that satisfy the equation. If the matrix A is real, then the complex roots occur in conjugate pairs. The roots are known as eigenvalues of A. Interesting questions include:
If lambda is an eigenvalue of A, then A-lambda*I is a singular matrix, and therefore there is at least one nonzero vector x with the property that (A-lambda*I)*x=0. This equation is usually written
A * x = lambda * xSuch a vector is called an eigenvector for the given eigenvalue. There may be as many as N linearly independent eigenvectors. In some cases, the eigenvectors are, or can be made, into a set of pairwise orthogonal vectors. Interesting questions about eigenvectors include:
Some useful facts about the eigenvalues lambda of a matrix A:
If a vector x is an exact eigenvector of a matrix A, then it is easy to determine the value of the associated eigenvalue: simply take the ratio of the I-th components of A*x and x, for any index I that you like.
But suppose x is only an approximate eigenvector, or worse yet, just a wild guess. We could still compute the ratio of corresponding components for some index, but the value we get will vary from component to component. We might even try computing all the ratios, and averaging them, or taking the ratios of the norms. However, the preferred method for estimating the value of an eigenvalue uses the Rayleigh quotient.
The Rayleigh quotient of a matrix A and vector x is
R(A,x) = (x^{T}*A*x) / (x^{T}*x).
The Rayleigh quotient is a scalar value whose magnitude generally lies between the smallest and largest magnitudes of the eigenvalues of the matrix. If the matrix A is symmetric, then all eigenvalues are real, and we may write the simpler statement:
Lambda_{min} <= R(A,x) <= Lambda_{max}
When x is an approximate eigenvector of A, the Rayleigh quotient is a very accurate estimate of the corresponding eigenvalue (if real) or the magnitude of the eigenvalue (if complex).
M File: write an M file called rayleigh.m which computes the Rayleigh quotient. It should have the form:
function r = rayleigh ( A, x )
Exercise: Copy the M file eigen_test.m from the web page. This file contains 5 test problems. The matrix you get by calling A=eigen_test(1) has eigenvalues 1/2, -1, and 2. Compute the value of the Rayleigh quotient for a variety of vectors x. Can you find a vector for which the quotient is exactly 2? More than 2? Less than -1?
x R(A,x) [ 1, 2, 3] -0.57 ____________________ __________ ____________________ __________ ____________________ __________ ____________________ __________ ____________________ __________What happens to the Rayleigh quotient when you look at a sequence of vectors x, A*x, A*A*x, ...?
The Rayleigh quotient R(A,x) will give us a way to approximate eigenvalues from an approximate eigenvector. Now we have to figure out a way to come up with an approximate eigenvector.
In many physical and engineering applications, the largest eigenvalue associated with a system represents the dominant and most interesting mode of behavior. For a bridge or support column, the eigenvalue might reveal the maximum load, and the eigenvector the shape of the object as it begins to fail under this load. For a concert hall, the dominant eigenvalue of the acoustic equation reveals the lowest resonating frequency. Hence, a simple means of approximating just one eigenvalue (the one of greatest magnitude) might be enough for some cases.
The power method is a simple iterative method for seeking the largest magnitude eigenvalue and its eigenvector. To discuss this method, let's begin with the matrix returned by A=eigen_test(1). I constructed this matrix to have a particularly simple structure. The eig command will show you the eigenvectors as well as the eigenvalues, if you use the command in the form:
[ V, lambda ] = eig ( A )The quantity V is actually a matrix, storing the N eigenvectors as columns. Record the eigenvalues and eigenvectors of A on this worksheet:
Eigenvalue: __________ __________ __________ Eigenvector Component 1: __________ __________ __________ Component 2: __________ __________ __________ Component 3: __________ __________ __________
The power method tries to determine the largest magnitude eigenvalue, and the corresponding eigenvector, of a matrix, by computing (scaled) vectors in the sequence:
x_{k+1} = A * x_{k}If we include the scaling, and an estimate of the eigenvalue, we get an algorithm that looks like this:
Starting with any nonzero vector x_old, repeat: Set x: = A * x_old Set r: = Rayleigh(A,x) Divide x by its L2 norm. If the maximum entry of x is negative, set x=-xThe last step is only done to make convergence easier to detect.
M File: write a simple M file that takes one step of the power method. Your file should have the form:
[x, r] = pow ( A, xold )
Exercise: Using your power method code, try to determine the dominant eigenvalue, and corresponding eigenvector, of the 5 eigen_test matrices. For consistency, use a starting vector of all 1's. Compare your results for matrix 1 with the results from the MATLAB eig command you recorded earlier.
Matrix Eigenvalue Eigenvector 1 __________ _____________________________ 2 __________ _____________________________ 3 __________ _____________________________ 4 __________ _____________________________ 5 __________ _____________________________One of these examples will probably give you slow convergence, and another may give you what seems like oscillatory behavior. Check the eigenvalues of these matrices. Check the norms of the eigenvalues. Can you make any observations?
The inverse power method reverses the iteration step of the power method. We write:
A * x_{k+1} = x_{k}or, equivalently,
x_{k+1} = A^{-1} * xIn other words, this looks like we are just doing a power method iteration, but now for the matrix A^{-1}. Leaving aside why we would want to do this, we can immediately assume that this process will tend to converge to the eigenvector associated with the largest eigenvalue of A^{-1}. Interestingly enough, the eigenvectors of A and A^{-1} are the same. Moreover, if v is an eigenvector of A for the eigenvalue LAMBDA, then it is an eigenvector of A^{-1} for 1/LAMBDA and vice versa.k
This means that we can freely choose to study the eigenvalue problem of either A or A^{-1}, and easily convert information from one problem to the other. Is there any advantage to looking at A^{-1}?
Well, the inverse power iteration will find us the biggest eigenvalue of A^{-1}, but that's the eigenvalue of A that's closest in magnitude to zero.
The method looks like this:
Starting with any nonzero vector x_old, repeat: Solve A*x: = x_old Set r: = Rayleigh(A,x) Divide x by its L2 norm. If the maximum entry of x is negative, set x=-x
M File: write a simple M file that takes one step of the inverse power method. Your file should have the form:
[x, r] = inpow ( A, xold )
Exercise: Using your inverse power method code, try to determine the smallest eigenvalue of the matrix eigen_test(5), starting from the vector of all 1's.
Step x R(A,x) 0 [ 1.00, 1.00, 1.00 ] 4.33 1 [ 0.87, 0.22, 0.44 ]_________ 3.00______ 2 _____________________________ __________ 3 _____________________________ __________ 4 _____________________________ __________ 5 _____________________________ __________
A major limitation of the power method was that it can only find the eigenvalue of largest magnitude. Similarly, the inverse power iteration seems only able to find the eigenvalue of smallest magnitude. But suppose we make what seems one tiny change. Instead of working with the matrix A, let's work with the matrix A+3*I, where we have shifted the matrix by adding 3 to every diagonal entry. How would this shift affect the inverse power method?
Well, we already said the method finds the eigenvalue of smallest magnitude. Denote the smallest eigenvalue of A+3*I by mu. But mu is an eigenvalue of A+3*I because A+3*I-mu*I is singular. But that means A-(mu-3)*I is singular. In other words, lambda=mu-3.0 is an eigenvalue of A, and is the eigenvalue of A that is closest to 3.
In general, the idea of shifting allows to focus the attention of the inverse power method on seeking the eigenvalue closest to any particular value we care to name. This also means that if we have an estimated eigenvalue, we can speed up convergence by using this estimate in the shift. A shifted inverse power method in which the shifts are adjusted on each step can be rapidly converging.
We are not going to worry here about the details of the expense of shifting and refactoring the matrix. (A real algorithm for a large problem would be very concerned about this!) Our simple-minded method looks like this:
Starting with any nonzero vector x_old and eigenvalue estimate r_old, repeat: Solve (A-r_old*I)*x: = x_old Set r: = r_old + Rayleigh(A-r_old*I,x) Divide x by its L2 norm. If the maximum entry of x is negative, set x=-x
M File: write a simple M file that takes one step of the shifted inverse power method. Your file should have the form:
[x, r] = inpows ( A, xold, rold )
Exercise: Using your shifted inverse power method code, we are going to search for the "middle" eigenvalue of matrix eigen_test(5). Assume we know that this mising eigenvalue must be within the range [ 1.5, 3.5], and start with a vector of all 1's. Start with a shift of 2.5, and thereafter, adjust your shift to the value of R(A,x).
Step x Shift or R(A,x) 0 [ 1.00, 1.00, 1.00 ] 2.5 1 [ 0.39, 0.91, -0.13 ]________ 3.34______ 2 _____________________________ __________ 3 _____________________________ __________ 4 _____________________________ __________ 5 _____________________________ __________Note that the cost of allowing the shift to vary is that we can't be sure that the code won't veer off to computing the largest or smallest eigenvalues. This could happen if the initial vector had a large component of the corresponding eigenvector. If you are interested, see if you can come up with a starting vector that causes your code to converge to the largest or smallest eigenvalue instead.
So far, the methods we have discussed seem suitable for finding one eigenvalue and eigenvector at a time. In order to find more eigenvalues, we'd have to do a great deal of special programming. Moreover, these methods are surely going to have trouble if the matrix has repeated eigenvalues, distinct eigenvalues of the same magnitude, or complex eigenvalues. The QR method is a more modern method which handles these sorts of problems in a uniform way, and computes all the eigenvalues and eigenvectors at one time.
The QR method repeats the following two steps:
The sequence of matrices A_{k} is orthogonally similar to the original matrix A, and hence has the same eigenvalues. Under fairly general conditions, the sequence of matrices A_{k} converges to
The MATLAB command
[ Q, R ] = qr ( A )computes the QR factorization (as does our own h_factor routine). The basic QR iteration is so simple that it's really not worth writing a routine to take a single step of it.
Exercise: Try out the QR iteration for the eigen_test matrices 1, 2 and 3. For matrix 2, which has 4 distinct eigenvalues of equal norm, you may find that the process does not behave as we would hope.
Serious implementations of the QR method save information about each step, which defines the eigenvectors. The algorithm also uses shifts to try to speed convergence. Some eigenvalues can be determined early in the iteration, which speeds up the process even more.
We are going to look at a simple version of the shifted QR method. To keep things distinct from the regular method, we'll call the matrix B here.
Let our initial matrix be B_{0}. Step k of the shifted QR method works as follows:
Assignment: compare 5 steps of the QR and shifted QR algorithms, starting with A_{0}=B_{0}=dif2(4), the difference matrix. For the shifted QR method, always use a shift of -1. At the end of 5 steps, record the 4 diagonal elements of your two matrices A_{5} and B_{5}.
diag(A5) diag(B5) __________ __________ __________ __________ __________ __________ __________ __________
Mail the 8 numbers to me.