Download and install
cvx. (We will experiment with this during the course to solve
Disciplined Convex Programming
Problems.)
(You can access MATLAB directly on the nexus machines or one of the
cpu servers such as cpu125.student.math or cpu09.student.math.
And also fe03.student.math, fe07.student.math, cpu03.student.math,
see the list)
Consider the MATLAB file:
closest_toeplitz_psd.m.
Run this file with the current data. Then,
remove the % to uncomment the two lines that generate a
random example. Run the file again.
Hand in your P,Z matrices and the error for each run.
CVX Problem: (Note: There is a typo in problem 4: $\alpha \geq
0$ is missing.)
consider the SDP relaxation of the max-cut problem discussed in class.
Solve the dual of the relaxation using CVX. Let y,X denote the
optimum of the dual and primal, respectively.
Generate random data as
input, i.e. generate random nonnegative weights for the graph and input
the Laplacian L as data for the SDP problem.
Find an (approximate) solution x for the original max-cut
problem and calculate a percentage error.
Hand in (or email) your cvx program and: L; primal dual optimal y, X;
the approximate x; and the percentage error.
Use CVX to solve the closest correlation
matrix problem, i.e. generate a random symmetric
n by n matrix X (try n=5,10,100) and
find a closest matrix in the intersection of: the cone of
positive semidefinite matrices and the subspace of symmetric matrices
with diagonal all ones.
Try and do the same for finding the nearest symmetric positive semidefinite
http://en.wikipedia.org/wiki/Toeplitz_matrix
Toeplitz matrix
, a symmetric matrix with constant 'diagonals'.
Check the accuracy of your solutions, e.g. is the optimum
you found positive semidefinite; are the primal and dual optimal values
equal?
(Hint for problem 1:
Use the following result from Linear Programming:
assume that $u,u^1,,\ldots,u^p$ are vectors in $\R^q$
If $u$ is a nonnegative combination of $u^1,\ldots, u^p$, then it must
be a nonnegative combination of no more than $q$ of these vectors.
(This follows from the definition of BFS in LP,
basic feasible solution in Linear Programming)
(Hint/Outline for problem 7:
\begin{enumerate}
\item
A convex function $f$ has bounded lower level sets if, and only if,
it satisfies the growth condition
\[
\liminf_{\|x\|\rightarrow \infty}
\frac {f(x)}{\|x\|} >0.
\]
\item
Since $f$ is closed, it is bounded below on compact sets.
\item
Therefore, the above inequality is equivalent to the existence of a
minorant of the form $\epsilon \| \cdot \|+k \leq f(\cdot)$ for some
constants $\epsilon >0$ and $k$.
\item
Taking conjugates, shows that this is equivalent to $f^*$ being
bounded above on a neighbourhood of $0$, which in turn is equivalent to
$f^*$ being continuous at $0$.
\end{enumerate}