\documentstyle{article}[12pt]
\pagestyle{empty}
\begin{document}
\newcommand{\bd}{\begin{description}}
\newcommand{\ed}{\end{description}}
\newcommand{\ds}{\displaystyle}
\newcommand{\vs}{\vspace{2 mm}}
\newcommand{\hs}{\hspace{1 mm}}
\newcommand{\sik}{\hs \hs \mbox{for} \hs \hs}
\newcommand{\nin}{\noindent}
\newcommand{\ol}{\overline}
\newcommand{\bc}{\begin{center}}
\newcommand{\ec}{\end{center}}
\newcommand{\be}{\begin{enumerate}}
\newcommand{\ee}{\end{enumerate}}
\newcommand{\up}{\vspace*{-1.5 in}}

\up
\bc
{\bf Practice problems for Exam 3}
\ec

\be

\item Let
\[ A= \left[ \begin{array}{ccc}
   1 & 1 & 0 \\
   0 & 2 & 0 \\
   0 & 1 & 1 \\
\end{array} \right]  \]
\be
\item Determine whether $A$ is diagonalizable.  If so, find a matrix
$S$
such that $S^{-1} A S$ is diagonal.  If not, explain why not.
\item What are the eigenvalues of $A^{-1}$?  Is $A^{-1}$
diagonalizable?  Explain.
\ee

\vs

(a) $A$ is indeed diagonalizable.  Computing the characteristic
polynomial, which is det$(A - \lambda I)$, one finds the eigenvalues to
be 1, 1, and 2.  By looking at the nullspace of $A - I$, and finding it
to be two-dimensional, one concludes that $A$ is indeed diagonalizable. 
So find three linearly independent eigenvectors and use them to construct
$S$.  And if this were an actual problem, you can check that you have a
valid diagonalizing matrix $S$ by computing the products $AS$ and $SD$
(where $D$ is a $3 \times 3$ matrix with the eigenvalues of $A$ on the
diagonal, in an order which conforms to your choice of ordering the
eigenvectors), and making sure they are the same \ldots because, you see,
$S^{-1} AS = D$ can only hold if $AS = SD$. 

\vs

(b) First one might ask how we know for sure that $A^{-1}$ even exists. 
Well \ldots in this case we know it from part (a), because $0$ is not an
eigenvalue of $A$, which is true if and only if $A$ is invertible.  Now
refer to Exercise 5.1.25, wherein one proves that if $A$ is invertible,
then the eigenvalues of $A^{-1}$ are the reciprocals of the eigenvalues
of $A$.  In the process of proving this, one finds that eigenvectors of
$A$ are also eigenvectors of $A^{-1}$.  In fact, the {\em eigenspaces} of
$A^{-1}$ and $A$ are the same \ldots which makes sense when you think
about this geometrically.  Consider $E_2$, for example:  The matrix $A$
doubles all vectors in this line, but these vectors don't change
direction.  So, considering that $A^{-1}$ ``undoes" what $A$ does, what
will $A^{-1}$ do to the vectors in this line?  It will ``halve" all of
them.  So the same line which is the eigenspace $E_2$ for $A$ is the
eigenspace $E_{1/2}$ for $A^{-1}$.  And the plane which remains fixed
under the action of $A$ will also remain fixed under the action of
$A^{-1}$. 

At any rate, $A^{-1}$ is indeed diagonalizable, and the eigenvalues are
1, 1, and 1/2. 

\vs
 
 
\item Produce an example of a $2 \times 2$ matrix with
eigenvalue 1, whose column space is the line $y = -x$ and whose
nullspace is the line $y = 2x$.


\vs

The statement of the problem tells you that the eigenvalues are 1 and 0. 
The fact that the nullspace is nontrivial means that your matrix can't be
invertible \ldots hence, 0 is an eigenvalue.  As we have two distinct
real eigenvalues, the matrix is diagonalizable.  So you should be able to
write down a factorized version of your matrix first, as $PDP^{-1}$, for
suitably constructed $P$ and $D$.  Then multiply out the product, and you
have a $2 \times 2$ matrix with the required properties. 

\vs

\item Suppose two animal species, slinkies and tinkies, coexist in
the same habitat.  Suppose $s_k$ is the slinkie population at time
$k$, and $t_k$ is the tinkie population at time $k$.  Let $\ds
\vec{x}_0 = \left[  \begin{array}{c}
                     s_0 \\ t_0   \\ \end{array} \right]$, and assume
the initial
populations $s_0$ and $t_0$ are positive.  Write down a dynamical
system describing this scenario, which predicts that if $t_0 < 2s_0$,
then both species will survive in the long run, but if $t_0 \geq
2s_0$, the slinkies will eventually die out.


\vs



I would approach this problem by starting with a drawing of a
``trajectory profile" \ldots something like the figure on page 349.  For
$2 \times 2$ dynamical systems with two real eigenvalues, both positive,
and neither equal to 1, we have three possible types of ``trajectory
profile"; the origin can be an attractor, a repellor, or a saddle.  So
which do you need here?  Start with a picture (which includes the whole
plane and not just the first quadrant) of the kinds of trajectories you
need.  The line $t = 2s$ will have to be a key feature of this picture. 
From the problem statement, we see that trajectories corresponding to
initial states {\em on} this line will have to head toward the origin
(read, both species are doomed).  So this line should be an eigenspace
corresponding to an eigenvalue between 0 and 1 (any such number will do;
this problem has many, many valid solutions).  The other eigenspace
should be another line.  Where should that line be?  (There are, in fact,
many lines which will work, but not just any slope will do.)  And what
sort of eigenvalue should the other one be?

Once you have chosen your eigenvalues and eigenspaces, then to get the
coefficient matrix for your dynamical system, you will have to write out
the full diagonalization factorization and then multiply \ldots much like
what you did in Problem 2.


\vs


\item Let $s$ and $t$ be numbers between 0 and 1.  Let
\[ A=\left[ \begin{array}{cc}
        s & 1-t \\
       1-s & t \\
\end{array}  \right]  \]
\be
\item Compute
\[ A \left[  \begin{array}{r}
                     1 \\ -1   \\ \end{array} \right]   \]
What does this computation tell you about $A$?
\item Explain why 1 is an eigenvalue of $A^T$. 
\item Compute $\displaystyle \lim_{k \to
\infty} A^k$.
{\em Hint:} Diagonalize $A$.
\ee
 

\vs



(a) It tells you that $s + t - 1$ is an eigenvalue of $A$, and that $\ds
\left[  \begin{array}{r}
                     1 \\ -1   \\ \end{array} \right]$ is a corresponding
eigenvector.

(b) This follows from a fact you may or may not have remembered (see
Exercise 5.1.29 and, while you're at it, Exercise 5.1.30):  If the rows
of a matrix all add up to the same number, then that number is an
eigenvalue (and in that case, notice that the vector of all ones is a
corresponding eigenvector).   

(c) For this part, you need to recognize that part (a) tells you that $s
+ t - 1$ is an eigenvalue of $A$, and that part (b) tells you that 1 is
also an eigenvalue of $A$ (because the eigenvalues of $A$ and $A^T$ are
the same.  You already have an eigenvector corresponding to $s + t - 1$;
to find an eigenvector corresponding to eigenvalue 1, look at the
nullspace of $A - I$.  I get the limiting matrix to be   
\[ \frac{1}{s + t - 2} \left[ \begin{array}{cc}
        t - 1 & t - 1 \\
       s - 1 & s - 1 \\
\end{array}  \right]   \]
(I hope that's correct.)

\vs


\item This is a three-part question; the three parts are related.
\be
\item Suppose $A$ is an $m \times n$ matrix with linearly independent
rows; suppose $\vec{y}, \vec{z} \in {\bf R}^m$, and suppose $A^T
\vec{y} = A^T \vec{z}$.  Show that $\vec{y} = \vec{z}$.
\item Suppose $A = BC$, where
\[ B = \left[  \begin{array}{rr}
            1/\sqrt{2} & 0 \\
             0 & 1/\sqrt{2} \\
            0 & -1/\sqrt{2} \\
            -1/\sqrt{2} & 0 \\
               \end{array}   \right] \hs \mbox{and} \hs
    C = \left[ \begin{array}{rrr}
               1 & -3 & 1 \\
               0   & 1 & 1 \\
\end{array}    \right]  \]
Show that if $A \vec{x} = \vec{b}$ is an inconsistent system, then
finding the least squares solution(s) to this system is equivalent to
solving the system $C \vec{x} = B^T \vec{b}$.
\item Let $A, B$, and $C$ be as in part (b).  Find all least squares
solutions to the system
\[  A \vec{x} = \left[  \begin{array}{c}
                     1 \\ 1 \\ 0 \\ 1 \\ \end{array} \right]  \]

\ee


\vs




(a) Rewrite $A^T \vec{y} = A^T \vec{z}$ as $A^T (\vec{y} - \vec{z}) =
\vec{0}$.  This is a homogeneous system, and since the columns of $A^T$
are linearly independent, the only solution is $\vec{0}$.  Therefore
$\vec{y} - \vec{z} = \vec{0}$. 

(b) Start with $A^T A \vec{x} = A^T \vec{b}$, and substitute $BC$ for
$A$.  Noting that $B$ has orthonormal columns and applying part (a)
appropriately, you can get the result.

(c) The point here is that you can use part (b), which enables you to
avoid multiplying $B$ and $C$ together. 


\vs


\item Let
\[ \vec{v_1} =
  \left[  \begin{array}{c}
                     4 \\ 1 \\ 1  \\ \end{array} \right], \hspace{4
mm}
\vec{v_2} =
  \left[  \begin{array}{c}
                    -1 \\ 3 \\  1 \\ \end{array} \right], \hspace{4
mm}
\vec{v_3} =
  \left[  \begin{array}{c}
                     -2 \\ -5 \\  13 \\ \end{array} \right]  \]
\be
\item Show that $\vec{v_1}, \vec{v_2}, \vec{v_3}$ form a
basis
for ${\bf R}^3$.
\item Give a formula for the linear transformation $S:
{\bf R}^3
\to {\bf R}^3$ which projects vectors onto the plane spanned by
$\vec{v_1}$
and $\vec{v_2}$.
\ee


\vs


 

(a) It's enough to verify that this is a mutually orthogonal triplet.

(b) I just want to point out here that the problem does not ask you for a
matrix; you are only asked for a formula, in which vectors are inputted,
and vectors are outputs.


\vs





\item Let $\ds A= \left[ \begin{array}{cc} 1/2 & 1/2 \\
1 & 0 \\ \end{array} \right]$.  Determine $\ds \lim_{n \to \infty}
A^n$.



\vs



This is somewhat like Problem 4, but simpler.  Diagonalize $A$ so that
you can take $n$th powers of the diagonal matrix, instead of $n$th powers
of $A$ itself.

\vs


\item Let $A$ be a $4 \times 4$ invertible matrix, and
suppose
$B$ is its inverse.
\be
\item Establish that the rows of $B$ are linearly independent. 
\item  Let $V = \hs \mbox{span} (\vec{B}^1, \vec{B}^2).$  What is the
dimension
of $V$?  Justify your answer.
\item  Find a basis for $V^\perp$.  Prove that the set of vectors you
choose
really is a basis for $V^\perp$.
\ee



\vs


 


(a) Well, $B$ is the inverse of $A$, so $B$ is also invertible.  And
matrices are invertible if and only if they have linearly independent
columns, and if and only if they have linearly independent rows.

(b) Taking all four rows of $B$ gives you a linearly independent set of
vectors; therefore, any subset of these is also a linearly independent
set.  Therefore the pair $\vec{B}^1, \vec{B}^2$ really is a basis for
$V$, and this gives $V$ dimension 2.

(c) So here's the tricky part of the problem.  You know that, since we're
working in ${\bf R}^4$, and since dim $V = 2$, the dimension of $V^\perp$
is also 2.  So you know you're looking for two vectors to form a basis. 
You need two vectors $\vec{y}$ and $\vec{z}$ such that $\vec{y}$ is
orthogonal to $\vec{B}^1$ and to $\vec{B}^2$, and $\vec{z}$ is orthogonal
to $\vec{B}^1$ and to $\vec{B}^2$.  Have we used the fact that $A$ and
$B$ are inverses?  Yes, but \ldots we only used it a little bit, in part
(a).  Think about the equation $BA = I$.  On the left is a matrix
product.  How are the individual entries of this matrix product
computed? 

\vs


\item Let
\[ V = \hs \mbox{span} \left( \left[ \begin{array}{r}  1 \\ 1\\ 0 \\
0 \\
\end{array} \right] , \left[ \begin{array}{r}  -1 \\ 1\\ 0 \\ 0 \\
\end{array} \right], \left[ \begin{array}{r} 0 \\ 0 \\ 1 \\ -1 \\
\end{array} \right] \right)   \]
Let $T: {\bf R}^4 \to {\bf R}^4$ be orthogonal projection onto $V$.
Find
the matrix $A$ which represents $T$.

\vs



First derive a formula like you did in 6(b).  Then remember that to
construct the matrix which represents $T$, you should apply $T$ to each
of the four standard unit vectors.

\vs



\item Let $V=$ span$(\vec{v}_1, \vec{v}_2, \vec{v}_3)$,
where
\[  \vec{v}_1 =  \left[ \begin{array}{r} 1 \\ -1 \\ 4 \\ 4 \\
\end{array}
\right] \hspace{.8 in}
   \vec{v}_2 =  \left[ \begin{array}{r} -1 \\ -3 \\ 0 \\ 8 \\
\end{array}
\right]  \hspace{.8 in}
   \vec{v}_3 =  \left[ \begin{array}{r} 2 \\ -16 \\ 3 \\ 1 \\
\end{array}
\right]  \]
Find an orthonormal basis for $V$.

\vs



This is just a straightforward application of the Gram-Schmidt algorithm,
but {\em remember, if you're really taking an exam, you should check that
the vectors you end up with really are mutually orthogonal}.  You don't
need to write that part down, but you should visually inspect your
vectors, and check mutual orthogonality in your head.  Don't forget to
normalize the vectors at the end; also remember normalizing is easier if
you scale to get rid of all fractions first.

\vs


\item Let
\[ A = \left[ \begin{array}{cccc}  1 & 2 & 3 & 4 \\ 0 & 1 & 2 & 3 \\
0 & 0 & 0 & 1 \\ 0 & 0 & 0 & 1 \\ \end{array}  \right]  \]
Find an orthonormal basis for the column space of $A$.

\vs

If you perform Gaussian elimination on $A$ (which requires one measly
step), you find the pivot columns to be $\vec{A}_1, \vec{A}_2$, and
$\vec{A}_4$.  So extract those three vectors (from the original matrix!)
as a basis, and apply Gram-Schmidt to get an alternate basis consisting
of mutually orthogonal vectors.  Don't forget to verify mutual
orthogonality and to normalize.

\ee
 
\end{document}
