MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsSpring 2011

Homework 1 Solutions

Exercise 1.1 a) Given square matrices A1 and A4 , we know that A is square as well:

A1 A2
A=
0 A4

=
Note that

det

I 0
0 A4

I 0
0 A4

A 1 A2
.
0
I

= det(I)det(A4 ) = det(A4 ),

which can be veried by recursively computing the principal minors. Also, by the elementary
operations of rows, we have

A1 0
A1 A2
det =
= det
= det(A1 ).
0
I
0 I
Finally note that when A and B are square, we have that det(AB) = det(A)det(B). Thus we have
det(A) = det(A1 )det(A4 ).
1
b) Assume A1
1 and A4 exist. Then

A1 A2
B1 B2
I 0
1
AA =
=
,
B3 B4
0 A4
0 I

which yields four matrix equations:

1. A1 B1 + A2 B3 = I,
2. A1 B2 + A2 B4 = 0,
3. A4 B3 = 0,
4. A4 B4 = I.
1
1
1
From Eqn (4), B4 = A
4 , with which Eqn (2) yields B2 = A1 A2 A4 . Also, from Eqn (3)
1
B3 = 0, with which from Eqn (1) B1 = A1 . Therefore,

1
A1
A1
A2 A41
1
1
.
A =
0
A1
4

Exercise 1.2 a)

0 I
I 0

A 1 A2
A 3 A4

A3 A4
A1 A2

b) Let us nd

B =

B1 B2
B3 B4

such that

BA =

A2
A1
1
0 A4 A3 A
1 A2

The above equation implies four equations for submatrices

1. B1 A1 + B2 A3 = A1 ,
2. B1 A2 + B2 A4 = A2 ,
3. B3 A1 + B4 A3 = 0,
4. B3 A2 + B4 A4 = A4 A3 A1
1 A2 .
First two equations yield B1 = I and B2 = 0. Express B3 from the third equation as B3 =
1
B4 A3 A1
1 and plug it into the fourth. After gathering the terms we get B4 A4 A3 A1 A2 =
A4 A3 A1
1 A2 , which turns into identity if we set B4 = I. Therefore

I
0
B =
A3 A1
I
1
c) Using linear operations
= 1. Then, det(A) = det(B)det(A) =

on rows we see that det (B)

A

A3 A1
.
Note
that
A
det (BA) = det (A1 ) det A4 A3 A1
2
4
1
1 A2 does not have to be invert
ible for the proof.
Exercise 1.3 We have to prove that det(I AB) = det(I BA).
Proof: Since I and I BA are square,

I
0
det(I BA) = det
B I BA

I A
I A
= det
B I
0 I

I A
I A
= det
det
,
B I
0 I
yet, from Exercise 1.1, we have

det

I A
0 I

= det(I)det(I) = 1.

Thus,

det(I BA) = det

2

I A
B I

Now,

det

I A
B I

= det

I AB 0
B
I

= det(I AB).

Therefore
det(I BA) = det(I AB).
Note that (I BA) is a q q matrix while (I AB) is a p p matrix. Thus, when one wants to
compute the determinant of (I AB) or (I BA), s/he can compare p and q to pick the product
(AB or BA) with the smaller size.
b) We have to show that (I AB)1 A = A(I BA)1 .
Proof: Assume that (I BA)1 and (I AB)1 exist. Then,
A = A I = A(I BA)(I BA)1
= (A ABA)(I BA)1
= (I AB)A(I BA)1
(I AB)1 A = A(I BA)1 .
This completes the proof.
Exercise 1.6
of limits, i.e.

a) The safest way to nd the (element-wise) derivative is by its denition in terms

d
A(t + t)B(t + t) A(t)B(t)
(A(t)B(t)) = lim
t0
dt
t
We substitute rst order Taylor series expansions
A(t + t) = A(t) + t

d
A(t) + o(t)
dt

B(t + t) = B(t) + t

d
B(t) + o(t)
dt

to obtain

d
1
d
d
(A(t)B(t)) =
A(t)B(t) + t A(t)B(t) + tA(t) B(t) + h.o.t. A(t)B(t) .
dt
t
dt
dt
Here h.o.t. stands for the terms

d
d
h.o.t. = A(t) + t A(t) o(t) + o(t) B(t) + t B(t) + o(t2 ),
dt
dt
a matrix quantity, where limt0 h.o.t./t = 0 (verify). Reducing the expression and taking the
limit, we obtain
d
d
d
[A(t)B(t)] = A(t)B(t) + A(t) B(t).
dt
dt
dt
b) For this part we write the identity A1 (t)A(t) = I. Taking the derivative on both sides, we have

d 1
d
d
A (t)A(t) = A1 (t)A(t) + A1 (t) A(t) = 0
dt
dt
dt
3

Rearranging and multiplying on the right by A1 (t), we obtain

d 1
d
A (t) = A1 (t) A(t)A1 (t).
dt
dt
Exercise 1.8 Let X = {g (x) = 0 + 1 x + 2 x2 + + M xM | i C}.
a) We have to show that the set B = {1, x, , xM } is a basis for X.
Proof :
1. First, lets show that elements in B are linearly independent. It is clear that each element in
B can not be written as a linear combination of each other. More formally,
c1 (1) + c1 (x) + + cM (xM ) = 0 i ci = 0.
Thus, elements of B are linearly independent.
2. Then, lets show that elements in B span the space X. Every polynomial of order less than
or equal to M looks like
M

p(x) =
i xi
i=0

for some set of i s.

Therefore, {1, x1 , , xM } span X.

b) T : X X and T (g(x)) =

d
dx g(x).

1. Show that T is linear.

Proof:

d
(ag1 (x) + bg2 (x))
dx
d
d
= a g 1 + b g2
dx
dx
= aT (g1 ) + bT (g2 ).

T (ag1 (x) + bg2 (x)) =

Thus, T is linear.
2. g(x) = 0 + 1 x + 2 x2 + + M xM , so
T (g(x)) = 1 + 22 x + + M M xM 1 .
Thus it can be written as follows:

0 1 0 0
0 0 2 0

0 0 0 3

.. .. .. .

. . . ..

0 0 0 0
0 0 0 0

.
.

0
0

0
1
2
3
..
.

0
M
4

1
22
33
.

M M
0

The big matrix, M , is a matrix representation of T with respect to basis B. The column
vector in the left is a representation of g(x) with respect to B. The column vector in the
right is T (g) with respect to basis B.
3. Since the matrix M is upper triangular with zeros along diagonal (in fact M is Hessenberg),
the eigenvalues are all 0;

i = 0 i = 1, , M + 1.

4. One eigenvector of M for 1 = 0 must satisfy M V1 = 1 V1 = 0

0


V1 = .

..

0
is one eigenvector. Since i s are not distinct, the eigenvectors are not necessarily inde
pendent. Thus in order to computer the M others, ones uses the generalized eigenvector
formula.

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsFall 2007

Homework 2 Solutions

Exercise 1.4 a) First dene all the spaces:

R(A) = {y Cm | x Cn such that y = Ax}
R (A) = {z Cm | y z = z y = 0, y R(A)}
R(A ) = {p Cn | v Cm such that p = A v}
N (A) = {x Cn | Ax = 0}
N (A ) = {q Cm | A q = 0}
i) Prove that R (A) = N (A ).
Proof: Let
z R (A) y z = 0 y R(A)
x A z = 0 x Cn
A z = 0 z N (A )
R (A) N (A ).
Now let
q N (A ) A q = 0
x A q = 0 x Cn
y q = 0 y R(A)
q R (A)
N (A ) R (A).
Therefore
R (A) = N (A ).
ii) Prove that N (A) = R(A ).

Proof: From i) we know that N (A) = R (A ) by switching A with A . That implies that

N (A) = {R (A )} = R(A ).
b) Show that rank(A) + rank(B) n rank(AB) min{rank(A), rank(B)}.

Proof: i) Show that rank(AB) min{rank(A), rank(B)}. It can be proved as follows:

Each column of AB is a combination of the columns of A, which implies that R(AB) R(A).

Hence, dim(R(AB)) dim(R(A)), or equivalently, rank(AB) rank(A).

Each row of AB is a combination of the rows of B rowspace (AB) rowspace (B), but the

dimension of rowspace = dimension of column space = rank, so that rank(AB) rank(B).

Therefore,
1

rank(AB) min{rank(A), rank(B)}.

ii) Show that rank(A) + rank(B) n rank(AB).
Let
rB = rank(B)
rA = rank(A)
where A Cmn , B Cnp .

Now, let {v1 , , vrB } be a basis set of R(B), and add n rB linearly independent vectors

{w1 , , wnrB } to this basis to span all of Cn , {v1 , v2 , , vn , w1 , , wnrB }.

Let
M=

v1 | v2

vrB | w1 |

wnrB

Suppose x Cn , then x = M for some Cn .

1. R(A) = R(AM ) = R([AV |AW ]).
Proof: i) Let x R(A). Then Ay = x for some y Cn . But y can be written as a linear
combination of the basis vectors of Cn , so y = M for some Cn .
Then, Ay = AM = x x R(AM ) R(A) R(AM ).
ii) Let x R(AM ). Then AM y = x for some y Cn . But M y = z Cn Az = x x
R(A) R(AM ) R(A).
Therefore, R(A) = R(AM ) = R([AV |AW ]).
2. R(AB) = R(AV ).
Proof: i) Let x R(AV ). Then AV y = x for some y CrB . Yet, V y = B for some Cp
since the columns of V and B span the same space. That implies that AV y = AB = x
x R(AB) R(AV ) R(AB).
ii) Let x R(AB). Then (AB)y = x for some y Cp . Yet, again By = V for some

CrB ABy = AV = x x R(AV ) R(AB) R(AV ).

Therefore, R(AV ) = R(AB).

Using fact 1, we see that the number of linearly independent columns of A is less than or equal to
the number of linearly independent columns of AV + the number of linearly independent columns
of AW , which means that
rank(A) rank(AV ) + rank(AW ).
Using fact 2, we see that
rank(AV ) = rank(AB) rank(A) rank(AB) + rank(AW ),
yet, there re only n rB columns in AW . Thus,
rank(AW ) n rB
rank(A) rank(AB) rank(AW ) n rB
rA (n rB ) rAB .
2

This completes the proof.

Exercise 2.2 (a) For the 2nd order polynomial p2 (t) = a0 + a1 t + a2 t2 , we have f (ti ) = p2 (ti ) +
ei i = 1, . . . , 16, and ti T. We can express the relationship between yi and the polynomial as
follows;

1 t1 t21

y1
e
1
a
. .
..
0
.

.

.
. .
.

a1 +

.

=
. .
.

1 t16 t16

a
2
e16
y16
The coecients a0 , a1 , and a2 are determined
by the least

squares solution to this (overconstrained)

problem, a = (A A)1 A y, where aLS = a0 a1 a2 .
Numerically, the values of the coecients are:

0.5296

aLS =
0.2061

0.375
For the 15th order polynomial, by a similar reasoning we can express the relation between data
points yi and the polynomial as follows:

1 t1 t21 t115
a0
e1
y1

.
.. ..
..
.
.
.

..
.

.
.

.
..

.

=

. .
2
15
y16
a15
e16
1 t16 t16 t16
This can be rewritten as y = Aa + e.
Observe that matrix A is invertible for distinct ti s. So

the coecients ai of the polynomial are aexact = A1 y, where aexact = a0 a1 a15 . The

resulting error in tting the data is e = 0, thus we have a perfect t at these particular time

instants.

Numerically, the values of the coecients of are:

0.49999998876521
0.39999826604650

0.16013119161635

0.04457531385982

0.00699544100513

0.00976690595462

0.02110628552919

0.02986537283027
aexact =

0.03799813521505

0.00337725219202

0.00252507772183

0.00072658523695

0.00021752221402

0.00009045014791

0.00015170733465

0.00001343734075

Part A Dashed: 2nddegree approximation ; Dashdotted: 15thdegree approximation

2.5

1.5

0.5

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Figure 2.2a
The function f (t) as well as the approximating polynomials p15 (t) and p2 (t) are plotted in
Figure 2.2b. Note that while both polynomials are a good t, the fteenth order polynomial is a
better approximation, as expected.
b) Now we have measurements aected by some noise. The corrupted data is
yi = f (ti ) + e(ti )

i = 1, . . . , 16

ti T

where the noise e(ti ) is generated by a command randn in Matlab.

Following the reasoning in part (a), we can express the relation between the noisy data points yi

and the polynomial as follows:

y = Aa + e
The solution procedure is the same as in part (a), with y replaced by y.
Numerically, the values of the coecients are:

0.00001497214861
0.00089442543781

0.01844588716755

0.14764397515270

0.63231582484352

1.62190727992829

2.61484909708492

2.67459894145774
aexact =
1.67594757924772

0.56666848864500

0.06211921500456

0.00219622725954

0.01911248745682

0.01085690854235

0.00207893294346
0.00010788458590
4

105

Part B Dashed: 2nddegree approximation ; Dashdotted: 15thdegree approximation

7

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Figure 2.2b

and

aLS

1.2239
= 0.1089
0.3219

The function f (t) as well as the approximating polynomials p15 (t) and p2 (t) are plotted in
Figure 2.2b. The second order polynomial does much better in this case as the fteenth order
polynomial ends up tting the noise. Overtting is a common problem encountered when trying
to t a nite data set corrupted by noise using a class of models that is too rich.
Additional Comments A stochastic derivation shows that the minimum variance unbiased
estimator for a is a
= argminy Aa2W where W = Rn1 , and Rn is the covariance matrix of the
random variable e. So,
a
= (A W A)1 A W y.
Roughly speaking, this is saying that measurements with more noise are given less weight in the
estimate of a. In our problem, Rn = I because the ei s are independent, zero mean and have unit
variance. That is, each of the measurments is equally noisy or treated as equally reliable.
c) p2 (t) can be written as
p2 (t) = a0 + a1 t + a2 t2 .
In order to minimize the approximation error in least square sense, the optimal p2 (t) must be such
that the error, f p2 , is orthogonal to the span of {1, t, t2 }:
< f p2 , 1 >= 0 < f, 1 >=< p2 , 1 >
< f p2 , t >= 0 < f, t >=< p2 , t >
< f p2 , t2 >= 0 < f, t2 >=< p2 , t2 > .

Part C Dashed: 2nddegree approximation

2.5

1.5

0.5

0.2

0.4

0.6

0.8

1.2

1.4

1.6

1.8

Figure 2.2c
We have that f = 12 e0.8t for t [0, 2], So,

< f, 1 >=
0
2

< f, t >=
0

< f, t2 >=

1 0.8t
5 8 5
e dt = e 5
2
8
8

t 0.8t
15 8 25
e dt = e 5 +
2
32
32

2 2
t

e0.8t dt =

85 8 125
e5
.
64
64

And,
8
< p2 , 1 >= 2a0 + 2a1 + a2
3
8
< p2 , t >= 2a0 + a1 + 4a2
3
8
32
< p2 , t2 >= a0 + 4a1 + a2
3
5
Therefore the problem reduces to solving another set of linear equations:

2 2 38
a0
< f, 1 >

2 8 4 a1 = < f, t > .

3
8
32
4
a2
< f, t2 >

3
5
Numerically, the values of the coecients are:

0.5353
a = 0.2032
0.3727
The function f (t) and the approximating polynomial p2 (t) are plotted in Figure 2.2c. Here
we use a dierent notion for the closeness of the approximating polynomial, p2 (t), to the original
function, f . Roughly speaking, in parts (a) and (b), the optimal polynomial will be the one for
6

which there is smallest discrepancy between f (ti ) and p2 (ti ) for all ti , i.e., the polynomial that will
come closest to passing through all the sample points, f (ti ). All that matters is the 16 sample
points, f (ti ). In this part however, all the points of f matter.

Exercise 2.3 Let y =

y1
y2

C1
e1
S1 0
,A=
,e=
and S =
.
C2
e2
0 S2

Note that A has full column rank because C1 has full column rank. Also note that S is symmetric
positive denite since both S1 and S2 are symmetric positive denite. Therefore, we know that
= (A SA)1 A Sy.
x
= argmin e Se exists and is unique and is given by x
Thus by direct substitution of terms, we have:
x
= (C1 S1 C1 + C2 S2 C2 )1 (C1 S1 y1 + C2 S2 y2 )
Recall that x
1 = (C1 S1 C1 )1 C1 S1 y1 and that x
2 = (C2 S2 C2 )1 C2 S2 y2 . Hence x
can be re-written
as:
x
= (Q1 + Q2 )1 (Q1 x
1 + Q2 x
2 )

Exercise 2.8 We can think of the two data sets as sequentially available data sets. x
is the
least squares solution to y Ax corresponding

to minimizing the euclidean norm of e1 = y Ax.
y
A
x
is the least squares solution to

x corresponding to minimizing e1 e1 + e2 Se2 where

z
D
e2 = z Dx and S is a symmetric (hermitian) positive denite matrix of weights.
By the recursion formula, we have:
x
=x
+ (A A + D SD)1 D S(z Dx
)
This can be re-written as:
x
=x
+ (I + (A A)1 D SD)1 (A A)1 D S(z Dx
)
)
=x
+ (A A)1 D (I + SD(A A)1 D )1 S(z Dx
This step follows from the result in Problem 1.3 (b). Hence
)
x
=x
+ (A A)1 D (SS 1 + SD(A A)1 D )1 S(z Dx
)
=x
+ (A A)1 D (S 1 + D(A A)1 D )1 S 1 S(z Dx
)
=x
+ (A A)1 D (S 1 + D(A A)1 D )1 (z Dx
In order to ensure that the constraint z = Dx is satised exactly, we need to penalize the
corresponding error term heavily (S ). Since D has full row rank, we know there exists at
least one value of x that satises equation z = Dx exactly. Hence the optimization problem we are
setting up does indeed have a solution. Taking the limiting case as S , hence as S 1 0, we
get the desired expression:
)
x
=x
+ (A A)1 D (D(A A)1 D )1 (z Dx

In the trivial case where D is a square (hence non-singular) matrix, the set of values of x over
in this
which we seek to minimize the cost function consists of a single element, D1 z. Thus, x
case is simply x
= D1 z. It is easy to verify that the expression we obtained does in fact reduce
to this when D is invertible.

Exercise 3.1 The rst and the third facts given in the problem are the keys to solve this problem,
in addition to the fact that:

R
UA =
.
0
Here note that R is a nonsingular, upper-triangular matrix so that it can be inverted. Now the
problem reduces to show that
x
= arg min y Ax22 = arg min(y Ax) (y Ax)
x

is indeed equal to
x
= R1 y1 .
Lets transform the problem into the familiar form. We introduce an error e such that
y = Ax + e,
and we would like to minimize e2 which is equivalent to minimizing yAx2 . Using the property
of an orthogonal matrix, we have that
e2 = U e2 .
Thus with e = y Ax, we have
e
22 =
U e22 = e U U e = (U (y Ax)) (U (y Ax)) = U y U Ax22

2
y1

= (y1 Rx) (y1 Rx) + y2 y2 .

y2
0

2
Since y2 22 = y2 y2 is just a constant, it does not play any role in this minimization.
Thus we would lik to have
y1 Rx
=0
and because R is an invertible matrix, x
= R1 y1 .

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsFall 2007

Homework 3 Solutions

Exercise 3.2 i) We would like to minimize the 2-norm of u, i.e., u22 . Since yn is given as
yn =

hi un1

i=1

we can rewrite this equality as

yn =

h1 h2

un1

un2
hn .

..

u0
We want to nd the u with the smallest 2-norm such that
y = Au.
where we assume that A has a full rank (i.e. hi =
0 for some i, 1 i n). Then, the solution
reduces to the familiar form:
u
= A (AA )1 y.

By noting that AA =
ni=1 h2i , we can obtain u
j as follows;
hj y
u
j =
n
2,
i=1 hi

for j = 0, 1, , n 1.

ii) a) Lets introduce e as an error such that yn = y e. It can also be written as y yn = e. Then
now the quantity we would like to minimize can be written as
r(y yn )2 + u20 + + u2n1
where r is a positive weighting parameter. The problem becomes to solve the following minimization
problem :
n

u
= arg min
u2i + re2 = arg min(u22 + re22 ),
u

i=1

from which we see that r is a weight that characterizes the tradeo between the size of the nal
error, y yn , and energy of the input signal, u.
In order to reduce the problem into the familiar form, i.e, y Ax, lets augment
is
bottom of u so that a new augmented vector, u

u
=

re
1

re at the

This choice of u
follows from the observation that this is the u
that would have u
22 = u22 + re2 ,
the quantity we aim to minimize .
Now we can write y as follows

y=

.
A ..

1
r

u
= Au
= Au + e = yn + e.

re

Now, u
can be obtained using the augmented A, A, as

A
1
1

u
= A (AA ) y = 1
AA +
y.
r
r
By noting that

1 2 1
AA + =
hi + ,
r
r

i=1

we can obtain u
j as follows
hj y
2
i=1 hi +

u
j = n

1
r

for j = 0, , n 1.

ii) b) When r = 0, it can be interpreted that the error can be anything, but we would like to
minimize the input energy. Thus we expect that the solution will have all the ui s to be zero. In
fact, the expression obtained in ii) a) will be zero as r 0. On the other hand, the other situation
is an interesting case. We put a weight of to the nal state error, then the expression from ii)
a) gives the same expression as in i) as r .

Exercise 3.3 This problem is similar to Example 3.4,

) = 0. We
t except now we require that p(T
can derive, from x(t) = p(t), that p(t) = x(t) tu(t) = 0 (t )x( )d where denotes convolution
and u(t) is the unit step, dened as 1 when t > 0 and 0 when t < 0. (One way to derive this
is to take x(t) = p(t) to the Laplace domain, taking into account initial conditions, to nd the
transfer function H(s) = P (s)/X(s), hence the impulse response, h(t) such that p(t) = x(t) h(t)).
t
T
Similarly, p(t) = x(t) u(t) = 0 x( )d . So, y = p(T ) = 0 (T )x( )d and 0 = p(T ) =
T
T
x(t) u(t) = 0 x( )d . You can check that < g(t), f (t) >= 0 g(t)f (t)d is an inner product on
the space of continuous functions on [0, T ], denoted by C[0, T ], which we are searching for x(t). So,
we have that y = p(T ) =< (T t), x(t) > and 0 = p(T ) =< 1, x(t) >. In matrix form,

y
< T t, x(t) >
=
= T t 1 , x(t)
0
< 1, x(t) >
where ., . denotes the Grammian, as dened in chapter 2. Now, in chapter 3, it was shown
= A A, A 1 y. So, for our problem,
that the minimum length solution to y = A, x , is x

1 y
x
= T t 1 T t 1 , T t 1
.
0
Where, using the denition of the Grammian, we have that:

< T t, T t > < T t, 1 >

T t 1 , T t 1 =
.
< 1, T t >
< 1, 1 >
2

No
use the denition for inner
the individual entries, < T t, T t >=
T w, we can
T product to nd
2
3
2
0 (T t) dt = T /3, < T t, 1 >= 0 (T t)dt = T /2, and < 1, 1 >= T . Plugging these in, one
can simplify the expression for x
and obtain x
(t) = 12y
[ 1 Tt ] for t [0, T ].
T2 2
Alternatively, we have that x(t) = p(t).
Integrating both
t sides and taking into account that
t t1
p(0) = 0 and p(0) = 0, we have p(t) = 0 0 x( )d dt1 = 0 f (t1 )dt1 . Now, we use the integration
t
t
t
by parts formula, 0 u dv = uv|t0 0 v du, with u = f (t1 ) = 0 1 x( )d, and dv = dt1 ; hence du =
tt
df (t1 ) = x(t1 )dt1 and v = t1 . Plugging in and simplifying we get that p(t) = 0 0 1 x( )d dt1 =
t
T
0 (t )x()d. Thus, y = p(T ) = 0 (T )x( )d =< T t, x(t) > . In addition, we have that
T
0 = p(T ) = 0 x( )d =< 1, x(t) > . That is, we seek to nd the minimum length x(t) such that
y = < T t, x(t) >
0 = < 1, x(t) > .
Recall that the minimum length solution x
(t) must be a linear combination of T t and 1, i.e.,
x
(t) = a1 (T t) + a2 . So,
y = < T t, a1 (T t) + a2 > = a1
0 =
< 1, a1 (T t) + a2 >
=

T
0

T
3
2
(T t)2 dt + a2 0 (T t)dt = a1 T3 + a2 T2
T
2
= a1 T2 + a2 T.
0 (a1 (T t) + a2 )dt

This is a system of two equations and two unknowns, which we can rewrite in matrix form:
T3 T2

y
a1
3
2
,
= T 2
a2
0
T
2
So,

a1
a2

T3
3
T2
2

T2
2

y
0

Exercise 4.1 Note that for any v C m , (show this!)

v v2

mv .

(1)

Therefore, for A C mn with x Cn

Ax2

Ax
2
mAx for x =
m x2 .
0, Ax
x2

But, from equation (1), we also know that

Ax2

x2

1
x

1
x2 .

Thus,

mAx
mAx
mA ,

x2
x

0, therefore
Equation (2) must hold for all x =
3

(2)

maxx=0

To prove the lower bound

Ax

1 A
n

Ax2
x2

= A2

mA .

A2 , reconsider equation (1):

Ax
Ax2
Ax2 for x = 0,

A2
x2
x2

But, from equation (1) for x C n ,

n
x2

Ax
x
for all x =
0 including x that makes

x1 . So,

nAx

nA2
x2

Ax
x

maxx=0

nAx
nAx2

nA2 .
x2
x2
(3)

maximum, so,

Ax
x

= A

nA2 ,

or equivalently,
1
A A2 .
n

Exercise 4.5 Any m n matrix A, it can be expressed as

0
A=U
V ,
0 0
where U and V are unitary matrices. The Moore-Penrose inverse, or pseudo-inverse of A,
denoted by A+ , is then dened as the n m matrix
1

0
A+ = V
U .
0
0
a) Now we have to show that A+ A and AA+ are symmetric, and that AA+ A = A and A+ AA+ =
A+ . Suppose that is a diagonal invertible matrix with the dimension of r r. Using the given
denitions as well as the fact that for a unitary matrix U , U U = U U = I, we have

AA

1 0
= U
V V
U
0
0

1
0

0
= U
I
U
0 0
0
0

Irr 0
= U
U ,
0
0

0
0 0

which is symmetric. Similarly,

+

1 0
0
0

1
0

Irr
0

A A = V
= V
= V

0
V
UU
0 0

0
0
V
I
0
0 0

0
V
0

which is again symmetric.

The facts derived above can be used to show the other two.

AA A =
=
=
=

Irr
(AA )A = U
0

Irr 0

UU
U
0
0

0
U
V
0 0
A.
+

0
0

U A

0
V
0 0

Also,

A AA

Irr 0
= (A A)A = V
V A+
0
0

1
Irr 0

= V
U
V V
0
0
0
0
1

0
= V
U
0
0
+

= A+ .
b) We have to show that when A has full column rank then A+ = (A A)1 A , and that when A
has full row rank then A+ = A (AA )1 . If A has full column rank, then we know that m n,
rank(A) = n, and

nn
A=U
V .
0
Also, as shown in chapter 2, when A has full column rank, (A A)1 exists. Hence

1
(A A) A =
V 0 UU
V
V 0 U
0

1
V 0 U
= V V

= V ( )1 V V 0 U

= V ( )1 0 U
= V ( 1 0 )U
= A+ .
5

Similarly, if A has full row rank, then n m, rank(A) = m, and

A = U mm 0 V .
It can be proved that when A has full row rank, (A A)1 exists. Hence,

A (AA )

= V
U U 0 V V
U
0

= V
U U U
0

= V
U U (1 )U
0
1

= V
U
0

= A+ .
c) Show that, of all x that minimize y Ax2 , the one with the smallest length x2 is given by
x
= A+ y. If A has full row rank, we have shown in chapter 3 that the solution with the smallest
length is given by
x
= A (AA )1 y,
and from part (b), A (AA )1 = A+ . Therefore
x
= A+ y.
Similary, it can be shown that the pseudo inverse is the solution for the case when a matrix A
has a full column rank (compare the results in chapter 2 with the expression you found in part (b)
for A+ when A has full column rank).
Now, lets consider the case when a matrix A is rank decient, i.e., rank(A) = r < min(m, n)
where A C mn and is thus neither full row or column rank. Suppose we have a singular value
decomposition of A as
A = U V ,
where U and V are unitary matrices. Then the norm we are minimizing is
Ax y = U V x y = U (V x U y) = z U y,
where z = V x, since is unaltered by the orthogonal transformation, U . Thus, x minimizes
Ax y if and only if z minimizes z c, where c = U y. Since the rank of A is r, the matrix
has the nonzero singular values 1 , 2 , , r in its diagonal entries. Then we can rewrite z c2
as follows:
z c2 =

r
n

(i zi ci )2 +
c2i .
i=1

i=r+1

It is clear that the minimum of the norm can be achived when zi = cii for i = 1, 2, , r and
the rest of the zi s can be chosen arbitrarily. Thus, there are innitely many solutions z and the

solution with the minimum norm can be achieved when zi = 0 for i = r + 1, r + 2, , n. Thus, we
can write this z as
z = 1 c,
where

1 =

1
0

r
0
0

and r is a square matrix with nonzero singular values in its diagonal in decreasing order. This
value of z also yields the value of x of minimal 2 norm since V is a unitary matrix.
Thus the solution to this problem is
x
= V z = V 1 c = V 1 U y = A+ y.
It can be easily shown that this choice of A+ satises all the conditions, or denitions, of pseudo
inverse in a).

Exercise 4.6. a) Suppose A Cn

m has full column rank. Then QR factorization for A can be

easily constructed from SVD:

n
A=U
V
0
where n is a nn diagonal matrix with singular values on the diagonal. Let Q = U and R = n V
and we get the QR factorization. Since Q is an orthogonal matrix, we can represent any Y Cpm
as

Y1
Y =Q
Y2
Next

Y1
R
Y1 RX
Y AXF2
= Q

2F
X 2F
= Q

Y2
0
Y2

Denote

D=

Y1 RX
Y2

and note that multiplication by an orthogonal matrix does not change Frobenius norm of the
matrix:
QD2F
= tr(D Q QD) = tr(D D) = D2F
Since Frobenius norm squared is equal to sum of squares of all elements, square of the Frobenius
norm of a block matrix is equal to sum of the squares of Frobenius norms of the blocks:

Y1 RX

2F
= Y1 RX F2 + Y2 2F

Y2

Since Y2 block can not be aected by choice of X matrix, the problem reduces to minimization of
Y1 RX2F . Recalling that R is invertible (because A has full column rank) the solution is
X = R1 Y1
7

b) Evaluate the expression with the pseudoinverse using the representations of A and Y from part
a):

AA

AY =

QQ

R
0

QQ

Y1
Y2

=R

Y1
Y2

= R1 Y1

From 4.5 b) we know that if a matrix has a full column rank, A+ = (A A)1 A , therefore both
expressions give the same solutions.
c)

Y
A
2
2
Y AXF + Z BXF =

X
F2
Z
B

A
Since A has full column rank,
also has full column rank, therefore we can apply results
B
from parts a) and b) to conclude that

X=

A
B

A
B

A
B

Y
Z

1
= A A + B B
A Y + BZ

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsSpring 2011

Homework 4 Solutions

Exercise 4.7 Given a complex square matrix A, the denition of the structured singular value
function is as follows.
(A) =

1
min {max ()| det(I A) = 0}

where is some set of matrices.

a) If = {I : C}, then det(I A) = det(I A). Here det(I A) = 0 implies that
there exists an x = 0 such that (I A)x = 0. Expanding the left hand side of the equation yields
1
x = Ax. Therefore 1 is an eigenvalue of A. Since max () = ||,
x = Ax
arg min{max ()| det(I A) = 0} = || = |

1
|.
max (A)

Therefore, (A) = |max (A)|.

0 such that
b) If = { Cnn }, then following a similar argument as in a), there exists an x =
(I A)x = 0. That implies that
x = Ax x2 = Ax2 2 Ax2
1
Ax2

max (A)
x2
2
1

max ().
max (A)
Then, we show that the lower bound can be achieved. Since = { Cnn }, we can choose
such that

1
max (A)

= V

..

U .

.
0

where U and V are from the SVD of A, A = U V . Note that this choice results in

1
0

V
=
V

I A = I V

0
1
which is singular, as required. Also from the construction of , max () =
(A) = max (A).

1
max (A) .

Therefore,

c) If = {diag(1 , , n )|i C} with D {diag(d1 , , dn )|di > 0}, we rst note that D1
exists. Thus:
det(I D1 AD) = det(I D1 AD)
= det((D1 D1 A)D)
= det(D1 D1 A)det(D)
= det(D1 (I A))det(D)
= det(D1 )det(I A)det(D)
= det(I A).
Where the rst equality follows because and D1 are diagonal and the last equality holds because
det(D1 ) = 1/det(D). Thus, (A) = (D1 AD).
Now lets show the left side inequality rst. Since 1 2 , 1 = {I| C} and 2 =
{diag(1 , , n )}, we have that
min {max ()| det(I A) = 0} min {max ()|det(I A) = 0},

which implies that

1 (A) 2 (A).
But from part (a), 1 (A) = (A), so,
(A) 2 (A).
Now we have to show the right side of inequality. Note that with 3 = { C}, we have 2 3 .
Thus by following a similar argument as above, we have
min {max ()| det(I A) = 0} min {max ()| det(I A) = 0}.

Hence,
2 (A) = 2 (D1 AD) 3 (D1 AD) = max (D1 AD).
Exercise 4.8 We are given a complex square matrix A with rank(A) = 1. According to the SVD
of A we can write A = uv where u, v are complex vectors of dimension n. To simplify computations
we are asked to minimize the Frobenius norm of in the denition of (A). So
(A) =

1
min { F | det(I A) = 0}

is the set of diagonal matrices with complex entries, = {diag(1 , , n )|i C}. Introduce
the column vector = (1 , , n )T and the row vector B = (u1 v1 , , un vn ), then the original
problem can be reformulated after some algebraic manipulations as
(A) =

1
minCn { 2 | B = 1}
2

To see this, we use the fact that A = uv , and (from excercise 1.3(a))
det(I A) = det(I uv )
= det(1 v u)
= 1 v u

Thus det(I A) = 0 implies that 1 v u = 0. Then we have

1 = v u

u1

.

.
.

v1
vn

un
n

v1 u1 vn un
..

=
.

n
= B

Hence, computing (A) reduces to a least square problem, i.e.,

min {F | det(I A) = 0} min 2 s.t. 1 = B.

We are dealing with a underdetermined system of equations and we are seeking a minimum norm
solution. Using the projection theorem, the optimal is given from o = B (BB )1 . Substituting
in the expression of the structured singular value function we obtain:

(A) =
|u

i vi |
2
i=1

In the second part of this exercise we dene to be the set of diagonal matrices with real entries,

= {diag(1 , , n )|i R}. The idea remains the same, we just have toalter the
constraint
Re(B)
equation, namely B = 1 + 0j. Equivalently one can write D = d where D =

and d =

Im(B)

1
. Again the optimal is obtained by use of the projection theorem and o = D (DDT )1 d.
0
Substituting in the expression of the structured singular value function we obtain:
1
(A) =

T
d (DDT )1 d

Exercise 5.1 Suppose that A C mn is perturbed by the matrix E C mn .

1. Show that
|max (A + E) max (A)| max (E).
3

Also nd an E that achieves the upper bound.

Note that

A = A + E E A = A + E E A + E + E A A + E E.
Also,
(A + E) = A + E A + E A + E A + E A E.
Thus, putting the two inequalities above together, we get that
|A + E A| E.
Note that the norm can be any matrix norm, thus the above inequality holds for the 2-induced
norms which gives us,
|max (A + E) max (A)| max (E).
A matrix E that achieves the upper bound is

E = U

1
0
..
.
0

0
..
.
0

0
.
..

V = A,

r
... 0

where U and V form the SVD of A. Here, A + E = 0, thus max (A + E) = 0, and

|0 + max (A)| = max (E)

is achieved.
2. Suppose that A has less than full column rank, i.e., the rank(A) < n, but A + E has full
column rank. Show that
min (A + E) max (E).

0 such that

Since A does not have full column rank, there exists x =

Ax = 0 (A+E)x = Ex (A+E)x2 = Ex2

(A + E)x2
Ex2
=
E 2 = max (E).
x2
x2

But,

min (A + E)
4

(A + E)x2
,

x2

as shown in chapter 4 (please refer to the proof in the lecture notes!). Thus
min (A + E) max (E).
Finally, a matrix E that results in A + E having full column rank and that achieves the upper
bound is

0
0
0

.
0 ..
0
0

.
..
. 0 r+1
..

E = U

0
r+1
0 0

V ,

for

A = U

0
...

0
..
.

0
..
.

V .

Note that A has rank r < n, but that A + E has rank n,

A + E = U

0
0
0
0

0
..
.
0
0
0

0
0
0
0
0
r
0
0 r+1
0
... r+1
0

V .

It is easy to see that min (A + E) = r+1 , and that max (E) = r+1 .
The result in part 2, and some extensions to it, give rise to the following procedure (which
is widely used in practice) for estimating the rank of an unknown matrix A from a known
matrix A + E, where E2 is known as well. Essentially, the SVD of A + E is computed, and
the rank of A is then estimated to be the number of singular values of A + E that are larger
than E2 .

Exercise 5.2 Using SVD, A can be decomposed as

..

A = U

V ,

where U and V are unitary matrices and k r + 1. Following the given procedure, lets select the
rst r+1 columns of V : {v1 , v2 , , vr+1 }. Since V is unitary, those vi s are orthonormal and hence
independent. Note that {v1 , v2 , , vr+1 , vn } span Rn , and if rank(E) = r, then exactly r of
the vectors, {v1 , v2 , , vr+1 , vn }, span R(E ) = N (E). The remaining vectors span N (E).
So, given any r + 1 linearly independent vectors in Rn , at least one must be in the nullspace of E.
That is there exists coecients ci for i = 1, , r + 1, not all zero, such that
E(c1 v1 + c2 v2 + cr+1 vr+1 ) = 0.
These coecients can be normalized to obtain a nonzero vector z, z2 = 1, given by

1
r+1

.

z=
i vi = v1 vr+1
.

.

i=1
r+1
and such that Ez = 0. Thus,

(A E)z = Az = U

v1
.

..

r+1

i vi = U

r+1 r+1

i=1
0

..

.
0

vr+1

1 1
2 2
..

(1)

By taking 2-norm of both sides of the above equation,

(A E)z2

1 1
2 2
.

1 1
2 2
.

..

= U r+1 r+1 2 = r+1 r+1

..

0
0

r+1

12
r+1

12

=
|i i |2
r+1
|i |2
.
i=1

( since U is a unitary matrix)

i=1

But, from our construction of z,

z22 = 1 ( v1
vr+1

1
1
r+1
.
2
..

2
|
i |
2 = 1.

=
1

=
)

2
2
.

i=1
r+1
r+1
6

(2)

Thus, equation(2) becomes

(A E)z2 r+1 .
Finally, (A E)z2 A E2 for all z such that z2 = 1. Hence
A E2 r+1
To show that the lower bound can be achieved, choose

..

E = U

V .

E has rank r,

A E = U

0
..

V .

.
0
r+1
..

.
k
0

and A E2 = r+1 .

Exercise 6.1 The model is linear one needs to note that the integration operator is a linear
operator. Formally one writes

e(ts) (u1 (s) + u2 (s))ds

0

(ts)
=
e
u1 (s) +
e(ts) u2 (s)

S(u1 + u2 )(t) =

= (Su1 )(t) + (Su2 )(t)

It is non-causal since future inputs are needed in order to determine the current value of y. Formally

one writes

e(ts) u(s)ds

(PT Su)(t) = (PT SPT u)(t) + PT

T

It is not memoryless since the current output depends on the integration of past inputs. It is also
time varying since

(ST u)(t) = (T Su)(t) +

e(tT s) u(s)ds

one can argue that if the only valid input signals are those where u(t) = 0 if t < 0 then the system
is time invariant.

Exercise 6.4(i) linear , time varying , causal , not memoryless

(ii) nonlinear (ane, tranlated linear) time varying , causal , not memoryless
(iii) nonlinear, time invariant , causal, memoryless
(iv) linear, time varying , causal, not memoryless
(i),(ii) can be called time invariant under the additional requirement that u(t) = 0 for t < 0

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsSpring 2011

Homework 5 Solutions

Exercise 7.2 a) Suppose c = 2. Then the impulse response of the system is

h(t) = 2(et e2t )

for t 0

One may assume that u(t) = 0 for t < 0 this will just alter the lower limit of integration in the
convolution formula, but will not aect the state space description, note also that the system is
causal.
t
y(t) =
2(e(t ) e2(t ) )u( )d t 0.
0

Hence by use of Leibniz dierentiation rule,

t
t
d (t )
2(t )
y(t) = 2
(e
e
)u( )d = 2 (2e2(t ) e(t ) )u( )d,
0 dt
0
and
t

y(t) = 2

(e(t ) 4e2(t ) )u( )d + 2u(t).

Now, let x1 (t) = y(t) and x2 (t) = y(t), then we have

x 1 (t)
x1 (t)
0
1
0
+
u(t).
=
2 3
2
x 2 (t)
x2 (t)
Since x1 (t) and x2 (t) can be written as x = Ax + Bu, there variables satisfy the continuous time
state property and are this valid state variables.
b) The transfer function of the system is
2(s + 2) c(s + 1)
, Re(s) > 1.
s2 + 3s + 2
When c = 2 there are no s terms in the numerator, which implies that the output y(t) only de
pends on u(t) but not on u (t). Our selection of state variables is valid only for c = 2. If c = 2, the
reachability canonical form may guide us to the selection of state variables
H(s) =

Exercise 7.3 In this problem, we have

y = a0 (t)y(t) + b0 (t)u(t) + b1 (t)u (t).
That is,

y=

a0 (t)y(t) + b0 (t)u(t) +
1

b1 (t)u (t).

Notice that
in the TI case, the coecients a0 , b0 , and b1 were constants, so we were able to integrate
the term b1 u (t). In this case, we can still get rid of the u (t), by integration by parts. We have

b1 (t)u (t)dt = b1 (t)u(t) u(t)b1 (t)dt.

So, our equation becomes

y = b1 (t)u(t) +

a0 (t)y(t) + (b0 (t) b1 (t))u(t).

Now, let
x = a0 (t)y(t) + (b0 (t) b1 (t))u(t),
we have that
y = x + b1 (t)u(t),
and substituting y in the equation for x we get:
x = a0 (t)x(t) + (b0 (t) b1 (t) a0 (t)b1 (t))u(t)
y = x + b1 (t)u(t)

Exercise 10.1 a) A =

J1

..
b) Let J
=
.

0 0
.
1 0

be the Jordan form decomposition of A, A = M JM 1 .

Jq
Note that
=0
= 0, 1 i q.
k
Also note that A = M J k M 1 and hence Ak = 0 J k = 0.
Thus, it suces to show that Jik = 0, for all i {1, . . . , q} for some nite positive power k i all
the eigenvalues of A are 0.
First, we prove suciency: If all the eigenvalues of A are 0, then the corresponding Jordan blocks

have zero diagonal elements and are such that Jini = 0 for every i, where ni is the size of Ji . Let

ko = max1iq ni . ko is nite and we have Jiko = 0, 1 i q.

Next, we proof necessity. Suppose there exists at least one eigenvalue of A, say io , that is non-zero.

Note that the diagonal elements of the k th power of the corresponding Jordan block(s) are kio , for

any positive power k. Hence, there exists at least one i such that Jik = 0, for any positive power k.

If A has size n, then the size of each of the Jordan blocks in its Jordan form decomposition is at

most n. Hence ko n and An = 0.

Jk

Jik

c) The smallest value is ko dened in part (b). Here, ko = nq .

J1

..
d) Let J =
be the Jordan form decomposition of A. We have that:
.
Jq
R(Ak+1 ) = R(Ak ) R(J k+1 ) = R(J k )
Thus it suces to look for the smallest value of k for which R(J k+1 ) = R(J k ). Note that a Jordan
block associated with a non-zero eigenvalue has full column rank, and retains full column rank
2

when raised to any positive power k. On the other hand, a nilpotent Jordan block of size ni has
column rank ni 1, and is such that rankJik = max{0, ni k}. Let N = {i|Ji is nilpotent}. Dene
kmin = maxiN ni . kmin is the smallest value of k for which R(J k+1 ) = R(J k ).

Exercise 11.1.
Since the characteristic polynomial of A is a determinant of a matrix zI A,
det(zI A) = det((zI A)T ) = det(zI AT ),
rst we show that
det(zI A1 ) = det(zI A2 ) = q(z)
for given A1 and A2 . For

qn1 1 0
qn2 0 1

.
. ..
A1 =

..
.. .

q1 0 0
q0 0 0

..
.

0
0
.

..
.

0
0

z
1 0

0
z
1

.
..
..
and zI A2 =
..
.
.

1
0 0
0
z
q 0 q1 q2

..
.

0
0
.

1
z + qn1

0
0
...

1
0
...

0
1
.
..

and A2 =

0
0
0

0
q0 q1 q2 qn1

we have

z + qn1 1 0
qn2
z 1

.
..
.
...

zI A1 =

.
.

0
0
q1
0
0

q0

..
.

0
0
.

..

Recall that det(A) = ai1 Ai1 + ai2 Ai2 + + ain Ain

a11 a12
a21 a22

A =
.
..

..
.

for

..
.

a1n
a2n
..
.

an1 an2

ann

where Aij is a cofactor matrix corresponding aij . Then,

det(zI A1 ) = (z + qn1 ) .
.
.
0

1 0

z
1

0
0

+ qn3 .
...
..

0
0

0
0

qn2 ..

0
z 1

0
0
z

0 0
1

0 0

0 0

.
.
q0 ..
..
.
.

.
.
.
.

0
z 1

0

0
z

1 0
z
1
0
z

..
.
..
..

.
0
0
0
0

0
0...
z

...

0
0

0
0
0
.
..

0
0
0

0 0
z
1
0
z

..
..
..
.

.
.

0 0
0 0

0
0
1 0
z
1
.
.
..
..

0
0
0
0

z 1
0
z

0
0
0
...

0
0
0

1 0

z
1

..
.

0
0
0
..
.

0
0
0
.

where the last depends on whether n is an even or odd number. Similarly if we take the
determinant of zI A2 using cofactors on the last row of zI A2 it is clear that we have
det(zI A1 ) = det(zI A2 ) = q(z).
Also it is true that
det(zI A) = det((zI A)T ) = det(zI AT ).
Hence
det(zI A1 ) = det(zI AT1 ) = det(zI A2 ) = det(zI AT2 ) = q(z).
Then we have
q(z) = (z + qn1 )z n1 + qn2 z n2 + + q1 z + q0
q(z ) = z n + qn1 z n1 + qn2 z n2 + + q1 z + q0 .
b) For A2 , we have

i 1 0
0 i 1

..
.
i I A2 = ...
.
..

0 0
0
q0 q1 q3

..
.

0
0
..
.

0
0
..
.

1
i
qn2 i + qn1

Suppose v1 is an eigenvector corresponding to i , then vi N (i I A2 ), i.e.,

0
0
0

i 1 0

0
i 1
0
0
0

0 0
i

0
0

v i = .

..
.
.

.
.
..

..
...

.
..
..
.

0 0

0
i
1
q0 q1 q3 qn2 i + qn1
0

(1)

It is clear from Eqn 1 that the only nonzero v i is

vi =

1
i
2i
..
.

n1
i

Then Eqn 1 becomes

i
0

q0

i
2
i
+
.

+
+

i q 1

0
..

n1
i

(i + qn1 )in1

0
..

q0 + q1 i +
+ ni 1 qn1 + ni

= 0

since i is a root of q().

c) Consider

0 1 0

A =

0 0 1

.

6 5 2
Its eigenvalues are 1 = 1, 2 = 3, and 3 = 2. Note that this A has the form of A2 thus the
corresponding eigenvectors can be written as follows:

1
1
1

v 1 =
1
,
v 2 =
2
,
v 3 =
3
.

12
22
23
Using those three eigenvectors we can obtain the
diagonal:

|
M =

v 1
|

similarity transformation matrix, M , to make A

|
|
v 2 v 3
.

|
|

Thus with

1 0 0
=

0 2 0
,

0 0 3
we have
A = M M 1 ,
which implies that


k1 0 0

= M k M 1 = M 0 k2 0 M 1

0 0 k

1
1 1
1
1 0 0
= 1 3 2 0 3 0 15
1
1
9 4
0
0 2
5

Ak

16
1
10

16
1
10
1
15

4
v15

and

e 1 t
0
0

0 M 1

= M 0 e2 t

0
0 e 3 t

1
1 1
e
0
0
1
= 1 3 2 0 e3t 0 15
1
e2t
1
9 4
0
0
5

eAt

61
1
10
4
15

16
1
10
1
15

Exercise 11.3 This equality can be shown in a number of ways. Here we will see two. One way
is by diagonalization of the matrix

.

Its characteristic equation is (A) = ( )2 + 2 , yielding eigenvalues = j. Using the
associated eigenvectors, we can show that it has diagonalization

1

1 1
j
j
2
2
=
.
1
1

j j
+ j
2 j 2
Now
jt

1 1
e e
exp t
=

e ejt
j j

jt
jt
jt
jt
e e +2e
e e e
j2
=
jt +ejt
jt
jt
e e j2
e e +e
2

e cos(t) e sin(t)
.
=
e sin(t) e cos(t)

1
2
1
2

j 12
j 12

An arguably simpler
way to

achieve this result is by applying the inverse Laplace transform

identity etA = L1 (sI A)1 . We have

s
(sI A) =

s
and so
1

(sI A)

1
=
(s )2 + 2
6

Taking the inverse Laplace transform element-wise gives us the previous result.
Exercise 11.4 This equality is shown through the denition of the matrix exponential. The
derivation is as follows.

1 k A
A
exp t
=
t
B
B
k!
k=0

1
tk Ak
=
tk B k
k!
k=0
1 k k
tA

t A
e
k=0
k!
=
=
1 k k
etB
k=0 k! t B
Exercise 11.5 By direct substitution for the proposed solution, we have:
etA BetA x(t) = etA BetA etA e(tto )(A+B) eto A x(to )
= etA Be(tto )(A+B) eto A x(to )

(2)

Dierentiating the proposed solution, we have:

dx(t)
= AetA e(tto )(A+B) eto A x(to ) + etA (A + B)e(tto )(A+B) eto A x(to )
dt
But since AetA = etA A, this is:
= etA (A + A + B)e(tto )(A+B) eto A x(to )
= etA Be(tto )(A+B) eto A x(to )

(3)

Since (3) and (2) are equal, the proposed solution satises the system of ODEs and hence is a
solution. Moreover, it can be shown that the solution is unique (though this is not the subject of
this class).

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsFall 2008

Homework 6 Solutions

Exercise 13.1 (a) The system described by

x = z
z = 4x3 + 2x
has an equilibrium point at (0, 0) for any value of (In particular, for
interval 1 1).
in
the

Additionally, when > 0 there are two more equilibrium points: 0, 2 (In particular, this is
the case for in the interval 0 < 1).
(b) Linearizing the system around (0, 0) we get the Jacobian:

0 1
A =
2 0
The characteristic polynomial of the system is det(A I) = 2 2. If > 0 there is an unstable
root, hence the linearized model is unstable around (0, 0). If < 0 both roots are on imaginary
axis, and the linearized system is marginally stable (neither asymtotically stable nor unstable). To
analyze stability of the original non-linear system in this case we would have to look at the higher
order terms.
For the two other equilibrium points (which exist for > 0) we get the Jacobian:

0
1
4 0
The characteristic polynomial for the system is det(A I) = 2 + 4. The complex conjugate
roots lie on the j axis and the linearized system is marginally stable.

Exercise 13.2 a) Notice that the input-output dierential equation can be written as
y = (u a1 y) + (u a2 y cy 2 )
and we can use observability-like realization employed for a discrete-time system of exercise 7.1
(c). The dierential equations for the states are
x1 = a1 x1 + x2 + u
x2 = a2 x1 cx21 + u
and the output equation is y = x1 . You can check that it is indeed a correct realization by
dierentiating the rst state equation and plugging in an expression for x2 from the second equation.
1

b) Let us consider the system with zero input and a1 = 3, a2 = 2 and c = 2.

x1 = 3x1 + x2
x2 = 2x1 2x21
The linearized systems matrix is

A =

3 1
2 0

with the characteristic equation 2 + 3 + 2 = 0, which has the roots 1 = 1 and 2 = 2.

Therefore the linearized system is asymptotically stable around the origin, which also means that
the original nonlinear system is a.s. around the origin (see lecture notes for the relevant theorem).
You can also verify by linearization that the other equilibrium point (1, 3) is unstable.
c) Let us nd a Lyapunov function for the linear system, and then nd a region where its derivative
is negative denite along the trajectories of the original nonlinear system. Since the linear system
is asymptotically stable, for any symmetric positive denite matrix Q there exists a unique positive
denite matrix P such that A P + P A = Q. Let us choose Q = I. Solving the system of linear
equations imposed by the matrix relation we obtain

P =

1
2

21
1

12

which gives us a quadratic Lyapunov function

1
V (x) = x P x = x21 x1 x2 + x22
2
Taking a derivative of V (x) along the trajectory using the chain rule we get:
V (x) = x21 x22 2x21 (2x2 x1 ) = x21 (1 + 2(2x2 x1 )) x22
The contour lines of the found Lyapunov function are
1 2
x x1 x2 + x22 = C 2
2 1

for various constants C. Let us nd such C that if the point x1 x2 is within the boundary then
V (x) < 0. Then the trajectory started in this set will stay there, and will asymptotically decay to
zero. Note that
1
1
V (x) = x21 + (x1 2x2 )2 = C 2
4
4
therefore
|2x2 x1 | < 2C
Therefore if C < 1/4 the derivative is strictly less than zero. Hence we found a region of attraction
as an ellipse, given by
1
x21 + (x1 2x2 )2 <
4
Any ball located completely within this ellipse will also be a region of attraction. Note also that
this set is not exhaustive, there are other points in space that converge to zero.

Exercise 14.2 (a) The system is asymptotically stable if all the roots of the characteristic poly
nomial lie in the left half of the complex plane. Note that characteristic polynomial for matrix A
in a control canonical form is given by
det(A I) = N + a0 N 1 + . . . + aN 1
b) Use continuity argument to prove that destabilizing perturbation with the smallest Frobenius
norm will place an eigenvalue of A + on the imaginary axis. Suppose that the minimum pertur
Assume that there is an eigenvalue in the right half plane. Consider a perturbation of
bation is .
where 0 c 1. As c changes from 0 to 1 at least one eigenvalue had to cross j axis,
the form c,
This proves contradiction
and the resulting perturbation has a smaller Frobenius norm than .

with the original assumption that A + has an eigenvalue in the right half plane.
c) The characteristic polynomial for the perturbed matrix is
det(A I) = N + (a0 + 0 ) N 1 + . . . + (aN 1 + N 1 )
We know that there exists a root = j, where is real. If we plug this solution in, and assemble
the real and imaginary parts and set them equal to zero, well get two linear equations in with
coecients dependent on ak and powers of . For example, for a 4th order polynomial:
(j )4 + (a0 + 0 )(j )3 + (a1 + 1 )(j)2 + (a2 + 2 )j + a3 + 3 = 0
results in the following two equations:
4 (a1 + 1 ) 2 + a3 + 3 = 0
(a0 + 0 ) 3 + (a2 + 2 ) = 0
This equation can be written in matrix form

0
2 0 1

3
0
0

as follows:

4 + a1 2 a3
1
=
2
a 0 3 a2
3

or
A() = B()
Therefore the problem can be formulated as nding a minimal norm solution to an underdetermined
system of equations:
min

A()=B()

By inspection we can see that matrix A has full row rank for any value of unequal to zero. If
= 0 the solution is 3 = a3 , and the rest of the k equal to zeros. For all other values of the
solution can be expressed as a function of :

() = A () A()A () 1 B()
Note that the matrix AA is diagonal, and can be easily inverted. By minimizing the norm of this
expression over we can nd
that corresponds to the minimizing perturbation. Then plug this

in the previous equation to compute minimizing perturbation (

)explicitly.
Compare
the
norms

of the solutions corresponding to =

and = 0 (i.e. compare (
) and a3 ), and choose the
3

minimum as the solution. This way we converted the problem to minimization of a function of a
single variable, which can be easily solved.
d) In case N = 2 the characteristic polynomial of the perturbed matrix is
2 + (a0 + 0 ) + (a1 + 1 ) = 0
where = j. For = 0 the minimizing solution is 1 = a1 , 0 = 0. If = 0, plug in = j,
and the resulting system of equations is
1 = 2 a1
0 = a0
This is a proper system (number of equations is equal to the number of unknowns), and its solution

is given directly by the equations. To minimize the norm of the solution we set = a1 . Note
that stability of original matrix A requires that a1 > 0, a0 > 0 (in fact positivity of all coecients
is always a necessary condition, but not sucient for N > 2 - use Routh criterion for a test in that
case!). Next, we have to compare |a1 | and |a0 |, and choose the smallest of them to null with 1
or 0 . In our problem a0 = a1 = a, therefore there are 2 solutions: (0, a) and (a, 0) for the set of s.

Exercise 14.5 a) Consider a nonlinear time-invariant system

x = f (x, u)
and its linearization around an equilibrium point (x,
u
) is
x = Ax + Bu.
1) Since the feedback u = Kx asymptotically stabilize the linearized system, all the eigenvalues

of the matrix A + BK are in OLHP.

2) Without loss of generality, we can take (x,

u
) = (0, 0) where the nonlinear system is linearized

around. Then,

x = f (x, u) =

f
f
x
+
u + (x, u),

u (x,
x (x,
u
)=(0,0)
u
)=(0,0)

where (x, u) is in the order of x2 .

If u = kx, the above equation can be written as follows:

f (x, u) = Ax + Bu + (x, kx),

where
(x, kx)
0 as x 0.
x
Thus the original nonlinear system is locally stabilized with the control law u = kx.
b) Consider the system S1 :
y + y 4 + y 2 u + y 3 = 0
where u is the input.

1) Let x1 = y and x2 = y, then state space representation of the system S1 is

x1
x2

x2
x31 x22 u x24

Thus the unique equilibrium point x is found to be x = (0, 0), which is independent on u .
2) Choose u = 0. Then the linearized system is

0
1
=
x
3x12 2x2 u 4x23

0 1
0
=
x +
u
0

0 0
x = Ax + Bu.

(x ,u )=(0,0)

x +

0
2

x2

(x ,u )=(0,0)

Since the eigenvalues of the matrix A are at 0, and the u term does not enter to the linearized
system, the linearized system can not conclude the local stability of the nonlinear system S1 .
c) Let u = c x32 where c is a function of x1 and x2 . Then
x 1 = x2

x 2 = x31 x22 (c x22 ) x42 = x31 cx22 .

So, it can be considered that this system has a nonlinear spring.

Now, choose a Lyapunov function candidate in the V (x) = x41 + x22 . Then,

V (x) =

4x31

2x2

x2
x31 cx22

= 4x31 x2 2x2 x13 cx32

= x2 (2x31 2cx22 )
= 2x2 (x31 cx22 ).
In order to make V (x) negative denite, we would like to have set
x21
x2
x21 + x22 + x31 x2
.
x32

x31 cx22 = x2
c =
This choice of c makes V (x) be

V (x) = 2x21 2x22 < 0 x =

0.
Thus the following control input

x21
u = 2x2 x2 +
x32
x2
stabilizes the system.
5

Exercise 14.7 (a) The linearized system is given by:

x 1
x 2

1
2x2
x2 (x1 + 1)

x1

x2

Hence, the matrix A of the linearized system is:

1 0
A=
0 1
A has repeated eigenvalues at 1. Thus, the nonlinear system is locally asymptotically stable about
the origin.
(b) The linearized system is given by:

x 1
x 2

3x21 1
1 1

x1

x2

Hence, the matrix A of the linearized system is:

0 1
A=
1 1

The characteristic polynomial is: (1 + ) 1 = 0. The eigenvalues are thus 1,2 = 12 25 (one
of the eigenvalues is in the right half plane), and the nonlinear system is unstable about the origin.
(c) The linearized system is given by:

x 1
x 2

1 1
2x1 1

x1

x2

Hence, the matrix A of the linearized system is:

1 1
A=
0 1
A has repeated eigenvalues at 1. Thus, the nonlinear system is locally asymptotically stable about
the origin.
(d) The linearized system is given by:

x1 (k + 1)
x2 (k + 1)

2 x2 (k)
1
1

x1 (k)

x2 (k)

Hence, the matrix A of the linearized system is:

2 0
A=
1 1
A has eigenvalues 1 = 2 and 2 = 1. Since one of the eigenvaues is outside the unit disk, the
nonlinear system is unstable about the origin.

(e) The linearized system is given by:

x1 (k + 1)
x2 (k + 1)

x2 (k)ex1 (k)x2 (k) x1 (k)ex1 (k)x2 (k)

1
2

x1 (k)

x2 (k)

Hence, the matrix A of the linearized system is:

0 0
A=
1 2
A has eigenvalues 1 = 0 and 2 = 2. Since one of the eigenvalues is outside the unit disk, the
nonlinear system is unstable about the origin.

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsSpring 2011

Homework 7 Solutions
Exercise 15.1 (a) The system is causal if the impulse response is right-sided. Consider a sequence
eat u[t], where u[t] is a unit step: u[t] = 1 for t 0, and zero otherwise. Laplace tranform of this
sequence converges if Re(s) > a, and is equal to

1
est eat u[t]dt =
, ROC : Re(s) > a
s
+
a

Therefore for a system represented by rst-order transfer function to be causal the ROC has to be
to the right of the pole (in fact this is true for a multiple pole as well). Since a rational function
can be represented by a partial fraction expansion, and region of convergence is dened by the
intersection of individual regions of convergence, the ROC of the system has to lie to the right
of the rightmost pole for the system to be causal. In case of a rational transfer function this is
also a sucient condition. Note that if an LTI system has a rational transfer function, its impulse
response consists of diverging or decaying exponents (maybe multiplied by powers of t), therefore all
concepts of p-stability are equivalent. For BIBO stability the impulse response has to be absolutely
integrable, which is equivalent to existence of Fourier transform. The Fourier transform is Laplace
transform evaluated at the j axis. Therefore for stability the ROC has to include the j axis.
Using these two rules we can see that the system
G(s) =

s+2
(s 2)(s + 1)

(i) is neither causal nor stable for ROC given by Re(s) < 1
(ii) is non-causal and stable for ROC 1 < Re(s) < 2.
(iii) is causal and unstable for ROC Re(s) > 2
Another way to solve the problem is to nd (look up in the tables) inverse Laplace tranforms
corresponding to the transfer function and ROC pairs. Compute partial fraction expansion

s+2
1
4
1
G(s) =
=

(s 2)(s + 1)
3 s2 s+1
The impulse response functions in three dierent cases are
(i) h(t) =
(ii) h(t) =
(iii) h(t) =

1
4e2t u[t] + et u[t] , Re(s) < 1 anticausal, unstable
3

1
4e2t u[t] + et u[t] , 1 < Re(s) < 2 non-causal, stable
3

1 2t
4e u[t] + et u[t] , Re(s) > 2 causal, unstable
3

(b) Note that if there is a diverging exponent in the impulse response, an input which is non-zero
on some interval will result in an exponentially diverging output. For example, in case (iii) choose
1

f (t) = 1 for 0 < t < 1 and 0 otherwise. The output for any positive t will be a linear combination
of e2t and et . For example for t > 1:

1
1 + 2(t )
2
y(t) =
4e
e(t ) u[t ]f ( )d = e2t 1 e2 et (e 1)
3
3
3
Clearly this function grows unbounded and has an innite p-norm. However the input f (t) has
p-norm equal to 1 for any p including . In case (i) we can use f (t) = 1 for 1 < t < 0, for
example. Innitely long bounded inputs that do not cancel an unstable pole will also result in
unbounded output. For example, choose e2t , t 0 in case (iii).

Exercise 15.2 a) When g(x) = cos(x), the system is unstable for p 1. Proof: Suppose the
system is p-stable. Then, there there exists a constant C such that gp Cup . Now, with
zp 0, there exists a T such that cos(z(t)) 1 for all t T . So, we have a contradiction
where up 0, then yp 1, which implies that there are no such a constantC to satisfy the
condition. Therefore the system is p-stable not all p 1.
b) When g(x) = sin(x), the system is p-stable for p 1. Proof: Consider the Taylor series
expansion of y(t) = sin(z(t)) about the origin. Then, we have
1
y(t) = sin(z(t)) = z(t) z 3 (t) + H.O.T.
3
This implies that
yp zp + O(zp ).

(1)

Now, because of the stability of the system from u to z , we have

zp Cup

(2)

for some constant C. Thus combining Eqn (1) and (2), we have
yp Cup + O(Cup ).
So, for all > 0 there exists such that O(Cup ) up , which implies that yp (C + )up .
That concludes the p-stability, with up < .
c) When g(x) is a saturation function with a scale of 1, the system is p-stable for p 1. Proof:
Again since the system from u to z is p-stable, there exists a constant C such that z p C up .
So, for all u with up , if we take C to be 1 , then we have:
zP Cup 1.
Since,

|g(z)| =

z
1

|z | 1
,
|z| 1

for |z| 1 we have

yp = zP Cup 1.
Therefore this system is p-stable for all p 1 in |z| < 1.

Exercise 16.1 a) Since u X, we can express u as

u=

ui eji t

where ui Rn , i R.

i=1

With
N

u (t) =

eji t =

i=1
N

u (t)u(t) = (

uTi eji t

i=1

eji t uTi )(

i=1
N
N

uj ej k t )

k=1

ej(k i )t uTi uk ,

i=1 k=1

we can compute Pu as follows:

Pu

L
1

= lim
u (t)u(t)dt
L 2L L
L

1
= lim
uTi
uj ej(j i )t dt
L 2L L
i
j
L

T
= lim
ej(j i )t dt.
u
i uj
L 2L
L
i

Note that as L , because of the orthnormality of complex exponential,

L
0 : i = j
lim
ej
(j i )tdt =
L L
1 : i=j
Thus
N

i=1

i=1

1
T
ui
ui (2L) =
ui 22 .
L 2L

Pu = lim

b) The output of the system can be expressed as y(t) = H(t) u(t) in time domain or Y (s) =
H(s)U (s) in frequency domain. For a CT LTI system, we have y = H(ji )ui eji t if u = ui eji t .
Thus
N

y(t) =
H(ji )ui eji t .
i=1

Following the similar method taken in a), we have

y (t)y(t) =

uTi H (ji )eji t

i=1
N
N

H(jk )uk ejk t

k=1

ej(k i )t uTi H (ji )H(jk )uk .

i=1 k=1

Thus, Py can be computed as follows:

Py

T
= lim
ui H (ji )H(jk )uk
L 2L
i
k
1

= lim
H(ji )ui 2 (2L).
L 2L

ej(k i )t dt

Py =

H(ji )ui 2 .

i=1

c)Using the fact shown in b),

Py =

i=1
N

H(ji )ui 2
2
max
(H(ji ))ui 2

i=1

=
Py
Py
sup Py =
Pu =1

2
max max
(H(ji ))
i

ui 2

i=1
2
max max (H(ji ))Pu
i
2
max max
(H(ji ))Pu
i
2
sup max
(H(j))Pu .

H2 .

d)Now we have to nd an input u X such that Py = H2 Pu .

1
|
|

.
.

H(j0 ) = U V =
u1
un

|
|
n

Consider

a SVD of H(j0 ):

v1

vn

Let u = v1 ej0 where 0 is such that H = max (H(j0 )), then

Py = H(j0 )v1 ej0 22 = H(j0 )v1 22 = 12 u1 22 .
2
(H(j0 )).
Py = max

Indeed, the equality can be achieved by the choice of u = v1 ej0 .

Exercise 16.3 We can restrict our attention to the SISO system since one can prove the MIMO
case with similar arguments and use of the SVD.
i.) Input l Output l
this case was treated in chapter 16.2 of the notes
ii.) Input l2 Output l2
this case was treated in chapter 16.3 of the notes
iii.) Input Power Output Power
this case was treated in the Exercise 16.1. Please note that Py = H2 Pu , the given entry in the
table corresponds to the rms values.
iv.) Input Power Output l2
a nite power input normally produces a nite power output (unless the gain of the system is zero
at all frequencies) and in that case the 2-norm of the output is innite.
v.) Input l2 Output Power
This is now the reveresed situation, but with the same reasoning. A nite energy input produces
nite energy output, which has zero power.
vi.) Input Power Output l
Here the idea is that a nite power input can produce a nite power output
whose -norm is
unbounded. Thinking along the lines of example 15.2 consider the signal u =

m=1 vm (t) where

3
vm (t) = m if m < t < m + m and otherwise 0. This signal has nite power and becomes un
bounded over time. Take that signal as the input to an LTI system that is just a constant gain.
vii.) Input l2 Output l

|y(t)
| =

h(t s)u(s) ds

|y(t)| = |< h(t s), u(s) >| h2 u2

The last step comes from the application of the Cauchy Schwartz inequality. Taking the sup on the
left hand side gives y h2 u2 ; now to achieve the bound apply the input u(t) = h(t)/h2 .
viii.) Input l Output l2
Apply a sinusoidal input of unit amplitude such that j is not a zero of the frequency response of
the transfer function H(j).
ix.) Input l Output Power
note that {u : u 1} is a subset of {u : Pu 1}. Therefore
sup{Py : u 1} sup{Py : Pu 1};
we use the lower bound from case iii.). Note that the entry in the table corresponds to rms and
not to power.

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsSpring 2011

Homework 8 Solutions
Exercise 17.4 1) First, in order for the closed loop system to be stable, the transfer function
from ( w1 w2 )T to ( y u )T has to be stable. The transfer function from w1 to y is given by
(I P K)1 P and is called system response function. The transfer function from w1 to u is given
by (I KP )1 and is called input sensitivity function. The transfer function from w2 to y is
(I P K)1 P K and is called the complementary sensitivity function. The transfer function from
w2 to u is given by (I KP )1 K. Therefore, we have the following :

y
(I + P K)1 P (I + P K)1 P K
w1
=
.
w2
u
(I + KP )1
(I + KP )1 K
So, if K is given as
K = Q(I P Q)1 = (I QP )1 Q,
then
(I + P K)1 P = (I + P Q(I P Q)1 )1 P
= (((I P Q) + P Q)(I P Q)1 )1 P
= (I P Q)P
(I + P K)

P K = (I P Q)P Q(I P Q)1

= P (I QP )(I QP )1 Q
= PQ
1

(I + KP )

= (I + (I QP )1 QP )1
= ((I QP + QP )(I QP )1 )1
= I QP

(I + KP )

K = (I QP )(I QP )1 Q
= Q.

Thus, the closed loop transfer function can be now written as follows:

y
(I P Q)P P Q
w1
=
.
u
(I QP )
Q
w2
In order for the closed loop system to be stable, then all the transfer functions in the large matrix
above must be stable as well.
(I P Q)P I P QP (I + P Q)P
(I + P Q)P P + P 2 Q
P Q P Q
I QP I + QP I + QP .
1

Since P and Q are stable from the assumptions, we know that all the transfer functions are stable.
Therefore the closed loop system is stable if K = Q(I P Q)1 = (I QP )1 Q.
2) From 1), we can express Q in terms of P and K in the following manner.
K = Q(I P Q)1
K(I P Q) = Q
K KP Q = Q
K = (I + KP )Q
Q = (I + KP )1 K = K(I + P K)1 ,
by push through rule.
For some stable Q, the closed loop is stable for a stable P . by the stabilizing controller K =
Q(I P Q)1 . Yet, not all stable Q can be used for this formulation because of the well-posedness
of the closed loop. In the state space descriptions of P and Q, in order for the interconnected
system, in this case K(s) to be well-posed, we have to have the condition ( 17.4 ) in the lecture
note, i.e., (I DP Q()) is invertible.
3) Suppose P is SISO, w1 is a step, and w2 = 0. Then, we have the following closed loop transfer
function:

1
Y (s)
(I P Q)P
=
,

U (s)
I QP
s

since the Laplace transform of the unit step is

1
s

we have

1
U (s) = (1 Q(s)P (s)) .
s
Then using the nal value theorem, in order to have the steady state value of u() to be zero, we
need:
1
u() = lim s(1 Q(s)P (s)) = 0
s0
s
1 Q(0)P (0) = 0
Q(0) = 1/P (0).
Therefore, Q(0) must be nonzero and is equal to 1/P (0). Note that this condition implies that P
cannot have a zero at s = 0 because then Q would have a pole at s = 0, which contradicts that Q
is stable.
Exercise 17.5 a) Let l(s) be the signal at the output of Q(s), then we have
l = Q(r (P P0 )l)
(I + Q(P P0 ))l = Qr
l = (I + Q(P P0 ))1 Qr.

Since we can write y = P l, and with P (s) =

from r to y can be calculated as follows:

2
s1 ,

P0 (s) =

1
s1 ,

and Q = 2, the transfer function

Y (s) = P (s)L(s)

Y (s)
R(s)

= P (I + Q(P P0 ))1 QR(s)

1

2
1
2
=
1+2

2R(s)
s1
s1
s1

4
s + 1 1
=
R(s)
s1 s1
4 s1
=
R(s)
s1 s+1
4
=
.
s+1

b) There is an unstable pole/zero cancellation so that the system is not internally stable.
c) Suppose P (s) = P0 (s) = H(s) for some H(s). Then using a part of the equation in a), we have
Y (s) = H(s)(I + Q(s)(H(s) H(s)))1 Q(s)R(s)
= H(S)I 1 Q(s)R(s)
= H(s)Q(s)R(s)

Y (s)
R(s)

= H(s)Q(s).

Therefore in order for the system to be internally stable for any Q(s), H(s) has to be stable.
Exercise 19.2 The characteristic polynom for the closed loop system is given by
s(s + 2)(s + a) + 1 = 0
Computing the locus of the closed poles as a function of a can be done numerically. The closed
loop system is stable if a 0.225. The above bound can also be derived by means of root locus
techniques or by evaluating the Routh Hurwitz criterion. Another way of deriving bounds for
the value of a is by casting this parametric uncertainty problem as an additive or multiplicative
perturbation problem, see also 19.5. One can expect that the derived bounds in such a case would
be rather conservative.
Exercise 19.4 We can represent an uncertainty in feedback conguration, as shown below.
Note that the plant is SISO, and we consider blocks and W to be SISO systems as well, so
we can commute them. The transfer function seen by the block can be derived as follows:

z = P0 (W w Kz)
= (I + P0 K)1 P0 W w
M

= (I + P0 K)1 P0 W.

Apply the small gain theorem, and obtain the condition for stability robustness of the closed loop
system as follows:

Figure 19.4

W (j)P0 (j)
< 1

sup

1 + P0 (j)K(j)

Exercise 19.5 a) Given

1
, K(s) = 10.
sa
In order for the system to remain stable, the zeros of 1 + P K be in open left half plane. Thus,
P (s) =

1 + PK = 1 +

10
s s + 10
=
a < 10.
sa
sa

b) Assume that the nominal plant is P0 = 1s . With W = a,

:

1
1
P0
s
=
=
,
1
1 + W P0
s a
1 a s

so when = 1, we have
1
P0
=
= P,
1 + W P0
sa
which says that P is clearly in .

c) The transfer function seen by the block was derived is (from the previous problem):

= (I + P0 K)1 P0 W.

Applying the small gain theorem:

M = sup |(1 + P0 K)1 P0 K | < 1

sup |

sup |

P0 W
|<1
1 + P0 K
1

j
10
+ j

| < 1

..
.
|a| <

2 + 100

|a| < 10.

Since block can have arbitrary phase we obtained much more conservative constraint on param
eter a than the one in a).
1
. With the same description of set, rst in
d) Now, the nominal plant is replaced by P0 = s+100
1
, we need to nd a new W as follows:
order to show that P with P0 = s+100
1

P0
1

s+100
=
=
,
1
1 + W P0
s
+
100
+ W
1 + W s+100
with = 1, the denominator becomes s + 100 + W , which we want to equate to s a. Thus we
have a new W to be
W = a 100.
Then in order to derive the condition on the closed loop system to be stable n the set , we use
the small gain theorem again.

sup |(1 + P0 K)1 P0 K| < 1

sup |

sup |

|a + 100|

<
..
.

P0 W
|<1
1 + P0 K
1

a100
j+100
|
10
+ j+100

<1

2 + 1102

210 < a < 10.

We can see that by representing uncertainty in a dierent way we can get a less conservative result.

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsSpring 2011

Homework 9 Solutions
Exercise 21.1 We can use additive perturbation model with matrices W and given in Figure 21.1.

W21
0
0 1
P11 0
K1 0
W =
,=
, P0 =
,K=
0
W12
2 0
0 P22
0 K2
Calculating transfer function from the output of block w to its input z we get

W21 K1 1
0
M = W12 K2 2 1+K11 P11
0
1+K22 P22
By assumption in the problem statement the decoupled system is stable, therefore the perturbed
system will be stable if I M does not have zeros in the closed RHP for any such that
1 1 and 2 1. By continuity argument this will be true if I M does not have
zeros on j axis, or equivalently |det (I M (j)(j))| > 0. Let us calculate the determinant in
question:

W21 K1 1
1
W12 W21 K1 K2 1 2
det (I M ) = det W12 K2 2 1+K11 P11 = 1
1
(1 + K11 P11 ) (1 + K22 P22 )
1+K22 P22
To have a stable perturbed system for arbitrary 1 1 and 2 1 it is necessary and
sucient to impose

W
W
K
K
12
21
1
2

(1 + K11 P11 ) (1 + K22 P22 )

< 1

Since uncertainty blocks enter as a product, the answer will not change if only one block is per
turbed.

Figure 21.1

Exercise 21.2
According to the lecture notes, we can state the problem equivalently as:
(M ) =

infRn {maxi | i | : i i ci = 1}

where ci = aibi and M = ab .

This problem is substantially more dicult because the determination of infRn {maxi | i | :
i i ci = 1} is equivalent to:

1

inf {max | i | : =
}
0
i
Rn
where:

R(c)
I(c)

From this point on, we proceed for

n = 3. The generelization to higher n is immediate.

1
Notice that the equality =
denes a line, say L0 R3 , in R3 . This makes our problem
0
equivalent to nding the smallest such that the cube B = { Rn : maxi | i | } touches the
line L0 . That can be done by looking at the following projections:
For every j such that I(cj ) = 0, do the following: Look for the smallest j such that maxi | i | =
j and

i (R(ci ) j I(ci )) = 1
i
R(cj )
I(cj ) . For each such j, the example done in the lecture notes
1

and the optimal is i = j sgn(R(ci ) j I(ci )).

i |R(ci )j I(ci )|

where j =

shows that the minimum

j is
By doing that, we found
the smallest j such that the projection of B j in j = 0 touches the projection of L0 . Among all
of the above candidate solutions, the only admissible is j such that j =
| j | j .
The nal solution is:

1
(M ) =
=
|R(ci ) j I(ci )|
j

I(ci ) i
i=j

cj

satises

Exercise 21.3 The perturbed system can be represented by the diagram in Figure 21.3. The closed
loop transfer function from reference input to output is
H(s) =

N (s)
D(s) + K(s)N (s)

The system is stable if the denominator does not have zeros in the closed RHP.

D + KN

D + KN = D0 + KN0 + D + KN = (D0 + KN0 ) 1 +

D0 + KN0
From the statement of the problem we know that K is a stabilizing controller, which means that
D0 + KN0 does not have zeros in the closed RHP. Therefore it is sucient to show that the second
parenthesis in the product above does not have zeros in the right half plane. By continuity argument

Figure 21.3
we can see that the minimum norm at which stability is lost puts at least one root on j axis.
Therefore we can rewrite the problem in the following way:
inf

min

D +KN

(j)=1
D0 +KN0

We can expand the constraint expression, taking into account that the real part has to be equal to
1 and imaginary part ot zero.
2
inf min

A =b

where

A =

D +KN
(j)
D

0 +KN0
D +KN
Im D0 +KN0 (j)

Re

, b =

1
0

The above represents underdetermined least squares problem. For all such that rank A() is 2,
the solution is
() = A (A A)1 b
The 2-norm of this expression can be minimized over , and compared to the solutions (if any)
with rank(A()) < 2.

Exercise 22.3 a) The modal test is the most convenient in this case. The system is reachable if
and only if rank [I A|B] = 5 (need to check for equal to the eigenvalues of A). Observe
that when = 2, [I A|B] is

0 1
b1

0
b2

0
b3

1 1
b4

1
b5
which has rank 5 if and only if b2 and b3 are linearly independent. Similarly, = 3, [I A|B] is

1 1
b1

1
b2

1
b
3

0 1
b4

0
b5
which has rank 5 if and only if b5 =
0.
(b) Suppose that A Rnn has k Jordan blocks of dimensions (number or rows) r1 , r2 , . . . , rk .
0. Furthermore, if blocks ri and rj have
Then we must have that br1 , br1 +r2 , . . . br1 +r2 +r3 ++rk =
the same eigenvalue, br1 +r2 +r3 ++ri and br1 +r2 +r3 ++rj must be linearly independent. These con
ditions imply that the input can excite the beginning of each Jordan chain, and hence has an impact
on each of the states.
(c) If the bi s are scalars, then they are linearly dependent (multiples of each other), so if two of
the Jordan blocks have the same eigenvalues the rank of [I A|B] is less than n.
Alternatively,
a) The system is reachable if none of the left eigenvectors of matrix A are orthogonal to B.
Notice that to control the states corresponding to a Jordan block, it is sucient to excite only
the state corresponding the beginning of the Jordan chain, or the last element in the Jordan block
(convince yourself of this considering a DT system for example). Thus it is not necessary that
generalized eigenvectors are not orthogonal to the B matrix! Besides, notice that if two or more
Jordan blocks have the same eigenvalue than any linear combination of eigenvectors corresponding
to those Jordan blocks is a left eigenvector again. In case (a) we can identify left eigenvectors of
matrix A:

0 1 0 0 0
w2 =

0 0 2 0 0
w3 =

0 0 0 0 1
w5 =

Any linear combination of w2 and w3 is also a left eigenvector. We can see that wk B = bk - k th row
of matrix B. Therefore for reachability of matrix A we need to have at least one non-zero element
in 5th row and linear independence of 2rd and 3rd rows of matrix B.

b) Generalizing to an arbitrary matrix in Jordan form we can see that all rows of matrix B corre
sponding to a Jordan block with unique eigenvalue should have at least one non-zero element, and
rows corresponding to Jordan blocks with repeated eigenvalues should be linearly independent.
c) If there are two or more Jordan blocks then we can nd a linear combination of the eigenvectors
which is orthogonal to the vector b, since two real numbers are obviously linearly dependent.

Exercise 22.4 The open loop system is reachable and has a closed-loop expression as follows:
xk+1 = Axk + B(wk + f (xk )),
where f () is an arbitrary but known function. Since the open loop system is reachable, there
exists the control input u such that
u =

u (0)

u (n 1)

that can drive the system to a target state in Rn , xf = x (n). Thus lets dene a trajectory
x (k) such that it starts from the origin and gets to x (n) by the control input u . Then, since
u(k) = w(k) + f (x(k)), let w(k) = u (k) f (x (K)). Then this w(k) can always take the system
state from the origin to any specied target state in no more than n steps.

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.

MASSACHUSETTS INSTITUTE OF TECHNOLOGY

Department of Electrical Engineering and Computer Science

6.241: Dynamic SystemsSpring 2011

Homework 10 Solutions
Exercise 23.1 a) We are given the single input LTI system:

0 1
0
x = Ax + bu , A =
, b=
0 0
1
The solution is expressed by:
t

At

eA(t ) bu( )d

x(t) = e x(0) +
0

Calculate exponent of matrix A by summing up the series and taking into account that An =
0 , n > 1.

1 t
At
e = I + At =
0 1
thus
e

At

b=

t
1

b) the reachability matrix is:

b Ab

0 1
1 0

The reachability matrix has rank 2, therefore the system is reachable. Now, we compute the
reachability Grammian over an interval of length 1:
1 1
1
A(T ) A(T )
G=
e
bb e
d = 31 2
0
2 1
The system is reachable thus the Grammian is invertible, so given any nal state xf we can always
nd such that xf = G. In particular

1
18
=
2 10
c) According to 23.5 dene F T (t) = eA(1t) b. Then u(t) = F (t) is a control input that produces
a trajectory that satises the terminal constraint xf . The control eort is given as:

u2 d = G

Infact this input corresponds to the minimum energy input required to reach xf in 1 second. This
can be veried by solving the corresponding underconstrained least squares problem by means of
the tools we learned at chapter 3.

d) First of all note that

G = xf G1 xf
The Grammian as well as its inverse are symmetric matrices. If we want to maximize the energy,
max{xf G1 xf | xf = 1}, we have to choose xf alligned with the singular vector corresponding to
min (G).

Exercise 23.4 Given :

x (t) = Ax + (b + )u,
where Rn , and (A, b) is reachable.
a) Using the Theorem 22.2, in order to make the system unreachable, we have wT B = for some
left eigenvectors wT of A. So, let i is an eigenvalue of A and wi be the corresponding left
eigenvectors. Then, using the theorem, we want to nd which makes this eigenmode unreachable
wiT (b + ) = 0. So, now we have
wiT = wiT b.
Then with this constraint, we would like to minimize 2 . Thus this can be cast into an optimiza
tion problem as follows:

Find
s.t.

min 2
wiT = wiT b.

This is exactly in the form of the least square problem. Since both and b are real, even when
wi Cn , let w
i = wiR wiI , where wiR and wiI are real and imaginary parts of wi respectively.
Then the formulation still remains as a least square problem as follows:

Find
s.t.

2
w
iT = w
iT b.

Then the solution to this problem is

= w
iT w
i )1 w
iT b
i (w

T
min 2 =
The last expression has to be minimized over all possible left eigenvectors of A. Note that the ex
pression does not depend on the norm of the eigenvectors, thus we can minimize over eigenvectors
with unity norm. If all Jordan blocks of matrix A have dierent eigenvalues, this is a minimization
over a nite set. In the other case we can represent eigenvectors corresponding to Jordan blocks
with the same eigenvalues as a linear combination of eigenvectors corresponding to particular Jor
dan blocks, and then minimize over the coecients in the linear combination.

b) NO. The explanation is as follows. With the control suggested, the closed loop dynamics is now
x = Ax + (b + )u
u = fT x + v
x = (A + (b + )f T )x + (b + )v.
Suppose that wi was the minimizing eigenvector of unity norm in part a). Then it is also an
eigenvector of matrix A + (b + )f T since wi is orthogonal to b + . Therefore feedback does not
improve reachability.

Exercise 24.5 a) The given system in general for all t 0 with u(k) = 0 k 0 has the following
expression for the output:
tk1
Bu(k)
y(t) =
k=1 CA
k1
= CAt
Bu(k)
k=1 A

since matrix A is stable. Note that because of stability of matrix A all of its eigenvalues are strictly
within unit circle, and from Jordan decomposition we can see that
lim Ak 2 = 0

therefore x() does not inuence x(0). Thus the above equation can be used in order to nd
x(0) as follows:
x(0) =

Ak1 Bu(k).

k=1

b) Since the system is reachable, any Rn can be achieved by some choice of an input of the
above form. Also, since the system is reachable, the reachability matrix R has full row rank. As
a consequence (RRT )1 exists. Thus, in order to minimize the input energy, we have to solve the
following familiar least square problem:
Find
s.t.

min u2

=
Ak1 Bu(k).
k=1

Then the solution can be written in terms of the reachability matrix as follows:
umin = RT (RRT )1 ,
so that its square can be expressed as
2
= uTmin umin
umin

= T (((RRT )1 )T RRT (RRT )1

= T (RRT )1 ,
3

where the last equality comes from the fact that inverse of a symmetric positive denite matrix is
still symmetric positive denite. Also, the Controlability Gramian of DT systems P is
P=

Ak BB T (AT )k = RRT ,

k=0

and is symmetric positive denite. Thus the square of the minimum energy, denoted as 1 (), can
be expressed as
1 () = T P 1 = M 22
where M is a Hermitian square root matrix of P 1 which is still symmetric positive denite.
c) Suppose some input umin results in x(0) = , then the output for t 0 can be expressed as
y(t) = Cx(t) = CAt .
Thus the square of the energy of the output for t 0 can be written as
y22 = (y T y)

C
C

CA
C A

=

..
..
.
.

= T
(AT )k C T CAk
T

t=0
T

= O O

Since the Observability Grammian of DT systems Q is

Q=

(AT )k C T CAk = OT O,
k=0

the square of the energy of the output for t 0 , which we now denote 2 (), can be expressed as
a function of as follows:
2 () T Q.
Also, because of the symmetric positive deniteness of Q, 2 () can be written as
2 () = N 22 ,
where N is a Hermitian square root matrix of Q.

d) It can be argued as follows:

= max{
y (t)2 |
u(t)2 1 , u(k) = 0k 0}
u

t=0

t=

= max{2 () | u s.t. = x(0) and

u(t)2 1 , u(k) = 0 , k 0}

t=

= max{2 ()

|
umin 22

1}

= max{2 () | 1 () 1}.

e) Now, using the fact shown in d) and noting the fact that P 1 = M T M where M is Hermitian
square root matrix which is invertible, we can compute :

= max{2 () | 1 () 1}

= max{N 22 | M 22 1} set = M 1 l

= max{(M 1 l)T OT OM 1 l | l22 1}

l

= max (OM 1 )
= max ((M 1 )T OT OM 1 )
= max ((M 1 )T QM 1 )
= max (QM 1 (M 1 )T )
= max (Q P)

Exercise 25.2 a) Given:

H1 (s) =

s+f
s + f
1
= 3
, H2 (s) =
.
3
2
s + 12s + 48s + 64
s2
(s + 4)

Thus the state-space realizations in controller canonical form for H1 (s) and H2 (s) are :

12 48 64
1

1
0
0
0 , C1 = 0 1 f
A1 =
, B1 =
, D1 = 0,
0
1
0
0
and
A2 = 2 , B2 = 1 , C2 = 1 , D2 = 0.
Since f is not included in the controllability matrix for H1 (s) with this realization, the controllabil
ity, which is equivalent to reachability for CT cases, the controllability is independent of the value
of f .Thus, check the rank of the controllability matrix:

1 12 96
1
12
rank(C) = rank 0
0
0
1
= 3.
Thus, the system with this realization is controllable. On the other hand, the observability matrix
O for H1 (s) contains f in it as follows:

0
1
f
1
f
0 .
O=
12 + f 48 60
Thus, when f = 4, O decreases its rank from 3 to 2.

Now, lets consider the state-space realization in observer canonical form for H1 (s). It can be

expressed as follows:

0 0 64
f

A1 = 1 0 48 , B1 = 1 , C1 = 0 0 1 , D1 = 0.
0 1 12
0
Since C1 does not contain f , the observability in independent of the value f . Thus check the rank
of the observability matrix:

0
0
1
1
12
rank(O) = rank 0
1 12 96
= 3.
Thus thus the system with this realization is observable.

On the other hand, the controllability matrix contains f in it as follows:

f 0 64
48 .
C= 1 f
0 1 f 12
Thus, again when f = 4, C decreases its rank from 3 to 2.

b) Let H(s) be the cascaded system, H2 (s)H1 (s). Then, the augmented system H(s) has the
following state-space representation:

A1
0
x1
B1
+
u
B2 C1 A2
x
0

x
1
0 C2
x2

12 48 64 0

0
0
0

0
u

x
+

0
0

1
0
0

0
1

f
2
0

0 0 0 1 x

x
2

=
y

x =

y =

x =

y =

Ax + Bu
Cx.

Here, we use A1 , B1 , and C1 from the controller canonical form obtained in a). Since matrix A has
zero block in its upper triangle, the eigenvalues of the cascaded system are ones of A1 and A2 , i.e.,
4, 4, 4, and 2. Thus the cascaded system is not asymptotically stable. Since C1 is not included
in the eigenvalue computation for A, the stability does not depend on the value of f .
The controllability matrix C for H(s) is

B AB A2 B A3 B

1 12 122 48 123 + 48 12 2 64
0
1
12
122 48
=

0
0

1
12
0
0
1
12 + f + 2

C =

which decreases its rank from 4 to 3 when f = 2. On the other hand, the observability matrix O
for H(s) is

CA

O =

CA2

C A3

0
0
0

0
1
f
=

1
f +2
2f
12 + f + 2 48 + 2f + 4 64 + 4f

2

,

thus the choice of f = 4, O drops its rank from full rank to 3. Thus the cascaded system is

unoberservable at f = 4.

It can be seen immediately that f = 2 case corresponds to unstable pole-zero cancellation. Thus,

for f = 2, the cascaded system is BIBO stable, but is not asymptotically stable due to the unstable

pole-zero cancellation.

MIT OpenCourseWare
http://ocw.mit.edu

6.241J / 16.338J Dynamic Systems and Control

Spring 2011

For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms.