This is an old revision of the document!
====== IAM 961 HW5 ====== ---- **Problem 1:** Implement Rayleigh quotient iteration and demonstrate its convergence, as follows. Develop your code with a small matrix (maybe 5 x 5) and do the final demonstration with a larger matrix (maybe 100 x 100). 1. Construct a random symmetric M x M matrix A with random but known eigenvalues and eigenvectors. Hint: construct A from a random orthogonal matrix V and a known random diagonal matrix D. (4 lines of code). 2. Select a eigenvalue of A and the corresponding eigenvector (an element of D and a column of V). Verify that these are in fact an eigenvalue, eigenvector pair. (3 lines of code). 3. To get an initial guess for the Rayleigh quotient iteration, perturb the eigenvector by 10% or so and evaluate the Rayleigh quotient using the perturbed eigenvector. (5 lines of code). 4. Now do the Rayleigh quotient iteration, stopping when $\|Av-\lambda v\| < 10^{-14}$. Plot the errors in the eigenvalue and eigenvector as a function of the iteration number. The Rayleigh quotient iteration code takes about 3 lines of code, the plotting more. Don't forget to label your axes! Answer these questions: Are you amazed, or what? Can you confirm that the errors scale as stated on page 208 of Trefethen and Bau? Note: the suggested line counts above are not requirements. They're meant as guidelines to let you know how roughly how much code each line should require. If you're writing more, you're working too hard. ---- **Problem 2:** Implement GMRES and investigate its convergence as a function of the matrix condition number, as follows. Given a value of $M$ and $\kappa$, write code that will 1. Construct a random M x M matrix A with condition number $\kappa$ and exponentially graded singular values, $\sigma_n = \kappa^{n/N}$. 2. Let $x$ be a random M vector with normally-distributed components. Let $b = A x$. 3. Do the GMRES iteration and plot the normalized residual $\| A x_n - b\|/\|b\|$ and the normalized $x$ error $\| x_n - x \|$ as a function of the iteration number $n$. Plot the residual with circles and the error with dots, put them on the same plot, and use a logarithmic axis. Set $M$ to a fairly small number, like 20 or 30 and produce three residual-and-error plots, for $\kappa=100$, $\kappa=10^6$, $\kappa=10^16$
