MIT 18.

06 Exam 3 Solutions, Fall 2017

Johnson

Problem 1 (30 points):

(a) Give a matrix A where det(A − λI) = 0 has roots λ = 1 and λ = 3, but
the trace of A does not equal 4.

(b) The eigenvalues of (A + AT )−1 for any real, square matrix A (assuming
A + AT is invertible) must be .
(c) If A = QT ΛQ for a diagonal matrix Λ and a real orthogonal matrix Q,
then the eigenvectors of A are the of Q.
     
2 1+i β
(d) If A is real, and eAt = e(3+4i)t +eαt , then the (t-
4 2 − 2i γ
independent) scalars α, β, γ are ?
d
(e) If A is a 4×4 matrix with det A = 5, then dt det(AT At) = .

Solution:
(a) The trace is the sum of the roots, including multiplicity, so all that we
need is a matrix where one of the roots is repeated. The easiest way to
do this is with a diagonal matrix, for example:
 
1 0 0
0 3 0
0 0 3

Here, the characteristic polynomial is (1 − λ)(3 − λ)2 , which has roots 1

and 3, but the trace is 7.

In hindsight, the question wording was somewhat ambiguous. You could

also read it as saying that λ = 1 and λ = 3 are among the roots, but are
not all of the roots. In that reading, you just need to provide a matrix
that has a third eigenvalue of any value. e.g. a diagonal matrix with 1, 2, 3
on the diagonal.

1
 (b) A + AT is real-symmetric, so (A + AT )−1 is real-symmetric as well. There-
fore its eigenvalues are real. As it is also invertible, none of its eigenvalues
are zero. Considering the 1 by 1 case, A = (a) for a 6= 0, we see that
(A + AT )−1 = 1/(2a) has eigenvalue 1/(2a), which can be any nonzero
real number depending on how we choose a. So that’s precisely the con-
dition: nonzero and real.
(c) A basis consisting of eigenvectors of A appears as the rows of Q, as it’s
written. The quick way to see this is that normally we write a diagonal-
ization as A = XΛX −1 where the columns of X are eigenvectors. Here,
since Q−1 = QT , we get the usual diagonalization with X = QT and
X −1 = Q. The columns of QT (the eigenvectors) are the rows of Q.
(d) A is real and (2, 4) is real, so if one eigenvalue is complex (here, 3+4i), the
solution must be a sum of complex-conjugate pairs. So, α = 3 − 4i, β = 1 − i, γ = 2 + 2i
.
(e) We have det(AT A) = det(AT ) det(A) = det(A)2 = 52 = 25 so det(AT At) =
t4 det(AT A) = 25t4 so

d d
det(AT At) = (25t4 ) = 100t3 .
dt dt

Problem 2 (30 points):

T
−B B
A=
You are given a matrix e for some real 3 × 3 matrix B. The nullspace
1
N (B) is spanned by  2 .
1

(a) Circle any of the following vectors that cannot possibly be eigenvectors
of A, and put a rectangle around any vectors that must be eigenvectors
of A:
             
1 1 0 1 1 2 1
 2  ,  0  ,  0  ,  0  ,  −1  ,  4  ,  1 
1 0 0 −1 1 2 1

(b) An x for some x 6= 0 may do what for large n (circle all possibilities)?
Oscillate / decay / diverge / go to a nonzero constant vector.
 
1
(c) For x =  0 , give a good approximation for An x for a very large n.
0

2
Solution:
First, let’s collect some observations. Recall from homework that N (B T B) =
N (B), so N (B T B) is also spanned by (1, 2, 1), and hence 0 appears as an eigen-
value of B T B with multiplicity 1. Remember also that B T B is symmetric
positive semidefinite, so −B T B is symmetric negative semidefinite, so it is diag-
onalizable and all of its eigenvalues are nonpositive real numbers. As we already
know that 0 appears as an eigenvalue with multiplicity 1, so all other eigenvalues
are strictly negative. Taking matrix exponentials exponentiates the eigenvalues
and preserves diagonalizability, eigenvectors, and multiplicity of eigenvalues.
T
So, putting all of this together, we see that e−B B is diagonalizable, has one
eigenvalue equal to 1 with corresponding eigenvector (1, 2, 1), and all other eigen-
values in the interval (0, 1). Because the matrix is symmetric, all eigenvectors
not on the line spanned by (1,2,1) must be orthogonal to it. So this gives the
answer to part (a): the (nonzero) vectors on the line spanned by (1,2,1) must
be eigenvectors, and any vector neither on that line nor perpendicular to it, or
that is 0, cannot be (one can also check that this is as much as you can possibly
say).
(a) The vectors that have to be eigenvectors (and should be rectangled) are
the multiples of the nullspace:
   
1 2
 2 ,  4 
1 2

and the vectors that cannot be eigenvectors (and should be circled) are
the ones that are not orthogonal to the above vectors (or are zero, which
cannot be an eigenvector):
     
1 0 1
 0 ,  0 ,  1 
0 0 1

(b) For any x, limn→∞ An x will exist and will equal the orthogonal projection
of x onto the line spanned by the vector (1, 2, 1) (because this line consists
of eigenvectors of eigenvalue 1, and all other eigenvalues are less than 1 in
absolute value. So in general this will allow the sequence An x to decay or
go to a nonzero constant vector , depending on whether x has a nonzero
component in the (1,2,1) direction.
(c) limn→∞ An x is the orthogonal projection of x onto the line spanned by

3
 (1, 2, 1). So this limit is:
  
1
1
 2  1 2 1  0   
1 0 1/6
lim An x =   =  2/6 
n→∞

1 1/6
1 2 1  2 
1

Problem 3 (40 points):

The vector x(t) satisfies the ODE

dx
(I + A) = (A2 − I)x
dt
 
0.9 0.0 0.3
for the diagonalizable matrix A =  0.0 0.8 0.4 . If we square this, we
  0.1 0.2 0.3
0.84 0.06 0.36
get A2 =  0.04 0.72 0.44  .
0.12 0.22 0.2

(a) If A has an eigenvalue λ and an eigenvector v, give a nonzero solution

x(t) satisfying the ODE above, in terms of λ, v, and t.
(b) Both A and A2 are matrices. By inspection
of A2 , what can you say (with no arithmetic! don’t calculate λ!) about
the magnitudes |λ| of the three eigenvalues of A2 ? What does this tell
you about the magnitudes |λ| of the eigenvalues of A?

 λof A with the biggest magnitude. A corresponding

(c) Give the eigenvalue
α
eigenvector is  2  for what α?
1
 
1
(d) For an initial conditions x(0) =  0 , circle what would you expect the
0
solutions x(t) to do for large t: Oscillate / decay / diverge / go to a
nonzero constant vector? Give a good approximation for x(t) for a
large t — if you can’t figure it out exactly, at least give a vector that x(t)
is nearly parallel to.

Solution:
(a) We can look for a solution pointing in the direction of v, i.e. a solution
of the form x(t) = c(t)v for some not-always-zero scalar valued function

4
 c(t). Notice that for this x(t), (I + A)x0 (t) = (I + A)c0 (t)v = (1 + λ)c0 (t)v
and (A2 − I)x(t) = (A2 − I)c(t)v = c(t)(A2 − I)v = (λ2 − 1)c(t)v. So, this
x(t) will satisfy the given ODE if and only if (1 + λ)c0 (t)v = (λ2 − 1)c(t)v
for all t. As A is invertible we must have λ 6= 0, so this means that we
need c0 (t) = (1 + λ)−1 (λ2 − 1)c(t), which simplifies because (λ2 − 1) =
(λ−1)(λ+1). We recognize a solution to this ODE given by c(t) = e(λ−1)t .
So this means that
x(t) = e(λ−1)t v
gives a solution as needed.

(Note that we’ll run into a problem if we have an eigenvalue λ = −1,

since then we would have 1 + λ = 0. Fortunately, in the next part we show
that this is not possible.)

(b) By inspection, both A and A2 are Markov matrices since the columns of
A clearly sum to 1. In fact A2 is a positive Markov matrix (positive en-
tries) so it has a 1-dimensional space of steady state vectors (= eigenvec-
tors with eigenvalue 1) and all other eigenvalues are less than 1 in absolute value .
As the eigenvalues of A2 , counted with multiplicity, are the squares of the
eigenvalues of A it follows that the same statements hold for the eigenval-
ues of A.

It turns out that this particular A turns out to have purely real, posi-
tive eigenvalues, but it’s not so easy to see that without more extensive
calculations..
(c) λ = 1 is the largest-magnitude eignevalue, from above. To find a corre-
sponding eigenvector, we should look for a vector in the nullspace of
 
−0.1 0.0 0.3
A − I =  0.0 −0.2 0.4  ,
0.1 0.2 −0.7
which is the same thing as looking for a nontrivial linear combination of
the columns that gives 0. If (x, y, z) is such a vector, the first row says
that x = 3z and the second row says that y = 2z, so we see that the vector
(3,2,1) is in the null space. So α = 3.
(d) Very similar to one of the homework problems, e(λ−1)t is a constant if
λ = 1 and is exponentially decaying for |λ| < 1 (where λ − 1 must
have a negative real part). So, if we expand the initial vector in the
basis of the eigenvectors (possible since A is given to be diagonalizable),
then x(t) is one constant term plus two decaying terms. So x(t) will


3
limit to a nonzero constant vector . This vector should be in the direction of  2  ,
1

5
 the eigenvector of λ = 1.

To determine the exact vector, similar to homework, we can easily show

that the sum of the components of x(t) is conserved. In particular, recall
that oT A = oT because A is a Markov matrix (here o is the vector with all
entires equal to 1). If we multiply both sides of the differential equation
by oT , we get
dx dx
oT (I + A) = 2oT = oT (A2 − I)x = (oT − oT )x = 0,
dt dt
so oT dx d T T
dt = dt (o x) = 0, hence o x is conserved. So the vector we’re
looking for is in the direction of (3, 2, 1) and has sum of coefficients equal
to the sum of coefficients of x(0) = (1, 0, 0), i.e. 1. It follows that
   
3/6 1/2
lim x(t) =  2/6  =  1/3  .
t→∞
1/6 1/6

Note that this problem could be simplified even further if you notice at the
beginning that (I + A)−1 (A2 − I) = (A − I), since A2 − I = (A + I)(A − I) =
A2 +IA−AI −I 2 , but you didn’t need this to solve the problem. (This approach
relies on I + A being nonsingular. Fortunately, from part b above, A cannot
have an eigenvalue −1 so I + A is indeed invertible.)