Chapter 2

Linear Algebra
Solution 2.1 Let the columns of A be denoted as a1 ; a2 ; a3 . Then a3 = a1 a2 . Hence the rank of A is 2
and 011 011 011
Im(A) = spanfa1 ; a2 g; Ker(A) = spanf@ 1 Ag; Im(A ) = spanf@ 1 A ; @ 0 Ag
1 0 1
Solution 2.2 D = orth(D0 ) and D? = null(D0)
0 0:4287 0:8060 1 0 0:4082 1
D = @ 0:5663 0:1124 A ; D? = @ 0:8165 A
0:7039 0:5812 0:4082
Solution 2.3 It is clear that (A) := max ja j satis es the norm conditions. Now let
1 
i;j ij

2
A = 3 4 ; 4 = (A) < (A) = 5:37
and 1   
A = 3 24 ; B = 11 11 ; 7 = (AB ) > (A) (B ) = 4:
Solution 2.4 Since the rank of A is 2, all solutions can be found by letting
 1 2  x   3   0 
4 1 x2 = 1 x3 + 1
1

and  x   1 2  1  3     0:7143   0:2857 

0
= 4 1 1 x3 + 1 = 1:8571 x3 + 0:1429
1
x2
so 0 x 1 0 0:7143 1 0 0:2857 1
@ x12 A = @ 1:8571 A + @ 0:1429 A
x3 1 0
for any . The minimal norm solution is given by
0 0:2181 1
x = A (AA ) 1 B = @ 0:0329 A
0:0947

 
Solution 2.5

x = (A A) 1 A B = 2:75

1
2 LINEAR ALGEBRA
Solution 2.6 Use MATLAB to get
A = V DV 1
where 0 0:7379 0:4243 0:8944 1
1 0 0 0
1
V = @ 0:4216 0:5657 0:4472 A ; D = @ 0 3 0A
0:5270 0:7071 0:0000 0 0 2
Then X (A) = spanfv2 ; v3 g and X+ (A) = spanfv1 g.
Solution 2.7 It is obvious that a1 = 1 P =1  n
i n
n
Q
and a = ( 1) =1  . It is also easy to see that det(I
n
i

A) =  (a11 + a22 + : : : + a ) +


+ a . Hence a1 = trace(A). Let  = 0 to get a = det( A) =

i i
n n
nn n n

( 1) det(A)
n

Solution 2.8 It follows easily from matrix inversion lemma

A
(A 1 + xy ) 1 = A Ax(1 + y Ax) 1 y A = A 1Axy
+ yAx :
Furthermore
det(A 1 + xy ) 1 = det A(I + xy A) 1 = det A det(I + xy A) 1
= det A (det(I + xy A)) 1 = det A (det(I + y Ax)) 1 = 1 +detyAAx :
Solution 2.9 Obvious.
Solution 2.10
R = 2Cnmax
k k=1
x
jx Axj  k k=1
; x
max k k=1
jx Ayj =  (A)
x ; y

Let x1 and 1 be such that Ax1 =  1x1 , kx1 k = 1, and (A) = j1 . Then
R  jx1 Ax1 j = j1 x1 x1 j = j1 j:
If A = A , then (A) = (A).
Solution 2.11
B (I + AB ) = (I + BA)B =) B (I + AB ) 1 = (I + BA) 1 B:
Solution 2.12 1. The eigenvalues of H are f2; 3; 2; 3g and the eigenvectors corresponding to 2 and
3 are the columns of 2 0 0:707
3
6 :5774 0:7071 77
V = 64 00:5774 0 5
0:5774 0
which form a basis for the stable invariant subspace.
2. The eigenvalues of H are f 1  j; 1; 1g and the eigenvector corresponding to 1 is
2 0:2887 3
6
V = 64 00::2887
77
2887 5
0:8860
which is a basis for the stable invariant subspace.
3. The eigenvalues of H are 0:4142j; 0:4142j; 2:4142j; 2:4142j and there is no stable invariant subspace.
 3

p=1,1.5,2,5,10, infinity
1.5

0.5

−0.5

−1

−1.5
−1.5 −1 −0.5 0 0.5 1 1.5

Figure 2.1: Visualizing the vector p-norm

Solution 2.13 The vector p-norm for p = 1; 1:5; 2; 5; 10; 1 are shown.
Solution 2.14 It is esay to verify that
I (I + A + A2 +


+ An )(I A) = An+1
Since kAk < 1, we have
I (I + A + A2 +


+ An )(I A) = An+1 ! 0; n ! 1:
Hence
I (I + A + A2 +


)(I A) = 0 or (I A) 1 = I + A + A2 +



Next
(I A) 1 = I + A + A2 +


 1 + kAk + kAk2 +


= 1 1kAk
Finally, note that
(I A)(I A) 1 = I
and
1 = (I A)(I A) 1  kI Ak (I A) 1  (1 + kAk) (I A) 1
Solution 2.15  Note that for any v 2 C k , we have
p
kvk1  kvk2  k kvk1
These inequalities imply that for any x 2 C n :
p1m kAxk2  kAxk1  kAxk2

=) p1 kkAx k2 1 kAxk2 kAxk1 kAxk2 kAxk2 p kAxk2

m xk2  pm kxk1  kxk1  kxk1  p1n kxk2 = n kxk2
p
=) p1 kAk  kAk  n kAk
m 2 1 2
4 LINEAR ALGEBRA
 Similarily, for any v 2 C k , we have
p
kvk2  kvk1  k kvk2
These inequalities imply that for any x 2 C n :
kAxk2  kAxk1  pm kAxk2
=) p1 kkAx k2  kAxk2  kAxk1  pm kAxk2  pm kAxk2
n xk2 kxk1 kxk1 kxk1 kxk2
p
=) p1 kAk2  kAk1  m kAk2
n
 Finally, for any v 2 C k , we have
kvk1  kvk1  k kvk1
These inequalities imply that for any x 2 C n :
kAxk1  kAxk1  m kAxk1
=) n1 kkAx k1  kAxk1  kAxk1  m kAxk1  m kAxk1
xk1 kxk1 kxk1 kxk1 kxk1
=) n1 kAk1  kAk1  m kAk1
Solution 2.16
kA B k22 = max [(A B ) (A B )] = max [(A A) (B  B )]
= max (A A)max (B  B ) = kAk22 kB k22 :
Solution 2.17
kxy k2 = max (xy (xy ) ) = py y max (xx )
p p

p p
= y y x x = kxk kyk :
Solution 2.18 The geometric interpolation of SVD:
M = U V 

U S V*

Figure 2.2: Geometric Interpolation of SVD

   
Solution 2.19 A = 0:55740:5574
1
2
0:5574j
0:5574 + 0:5574j
1:1547 and B = 1 j . 1
 5
 
See Horn and Johnson. Let P = P  = P11 P12  0 with P 2 C kk . Show  (P ) 
Solution 2.20
P12 P22 11 i
i (P11 ); 8 1  i  k.
P p;m)  2  min(p; m) 2 .
Solution 2.21 Let R = U V  . Then trace R R = trace V 2 V  = trace 2 = min(
i=1 i 1
p
kRkF  min(p; m)  max (R):
Solution 2.22  Note that for a suciently small  > 0
8 9
< X =
kI + Ak1 kI k1 = max
i :
1 + Aii + jAij j; 1
j =1;j 6=i
8 9
< X =
=  max A +
i : ii
jAij j;
j =1;j 6=i
so 8 9
lim+ kI + Ak1 kI k1 = max <A + X =
1 (A) = ! 0  i : ii
jAij j;
j =1;j 6=i

8 9
< X =
kI + Ak1 kI k1 = max
j :
1 + Ajj + jAij j; 1
i=1;i6=j
8 9
< X =
=  max A +
j : jj
jAij j;
i=1;i6=j
so 8 9
lim+ kI + Ak1 kI k1 = max
< X =
1 (A) = ! 0
A +
i : jj
jAij j;
i=1;i6=j

p p
kI + Ak2 kI k2 = max (I + A) (I + A) 1 = max (I + (A + A) + 2 A A) 1
n o
1  1 + 2 max (A + A) 1 = 2 max (A + A)
p
 1 + max (A + A)
so

lim+ kI + Ak2 kI k2 = max ( A 2+ A )
2 (A) = ! 0
Solution 2.23 Note that
     
X I = XI 1 I0 X 0 I X 1
I Y 0 Y X 1 0 I
so  
X I  0 () Y X 1  0 () min (X 1=2 Y X 1=2)  1 () min (XY )  1:
I Y
6 LINEAR ALGEBRA
Solution 2.24
01 2 3
1
P = @2 5 5 A = V DV 
3 5 10
where 0 0:9623 0:0596 0:2656 0
1 0 0
0 1
V = @ 0:1925 0:8389 0:5092 A ; D = @ 0 1:9172 0 A
0:1925 0:5410 0:8187 0 0 14:0828
Hence 0 0:2696 0:5766 0:7713 1
P 1 2 = V D1 2 V  = @ 0:5766 1:9473 0:9358 A
= =

0:7713 0:9358 2:9205

Solution 2.25
09 19 29
1 0
1 2 1 0
1  0 1 2 1
P = @ 19 A @
41 63 = 3 4 A 0 2 @3 4A
29 63 97 5 6 5 6
So  
p1
C= 2 2 4 2 6 2 p3 p5
Solution 2.26
2A A A 3
= min 4 X111 A1222 X133 5
X1 ;X2 ;X3
X2 A32 A33
 
= min max k A11 A12 A13 k ; AA22 X3
32 A33
 
X3

= max k A11 A12 A13 k ; AA22 ; k A32 A33 k

32

The minimizing X3 is obtained from

min A22
A A
X3
X3 32 33

With the above X3 , then X1 and X2 are obtained from

2A A A 3
= min 4 X111 A1222 X133 5
X1 ;X2
X2 A32 A33
Solution 2.27 p
= jaj2 + jbj2 and x = jbbajc2 .
0
Chapter 3

Linear Dynamical Systems

Solution 3.1

X_ (t) = dtd eAtX (0)eBt = AeAt X (0)eBt + eAt X (0)eBtB = AX (t) + X (t)B:

Solution 3.2 ???? where

_ t; t0 ) = A(t)(t; t0 ); (t0 ; t0 ) = I:
(
(Show also that  1 (t;  ) = (; t) and (t; )(;  ) = (t;  ).)
Note that
0 = dtd  1 (t; t0 ) (t; t0 ) = d dt(t; t0 )  +  1 (t; t0 ) d(dtt; t0 ) ; ) d dt(t; t0 ) =  1 A

1 1

and mmultiply  1 (t; t0 ) from the left to the di erential equation to get
 1 (t; t0 ) (x_ A(t)x(t)) =  1 (t; t0 )B (t)u(t)
i.e.,
d  1 (t; t )x(t)
=  1 (t; t )B (t)u(t)
dt 0 0

Integrate both sides to get

Z t
 1 (t; t0 )x(t)  1 (t0 ; t0 )x(t0 ) =  1 (; t0 )B ( )u( )d
t0
i.e. Z t
x(t) = (t; t0 )x(t0 ) +  1 (t; t0 ) 1 (; t0 )B ( )u( )d
t0
The result follows by noting
 1 (t; t0 ) 1 (; t0 ) = (t;  ):
Solution 3.3
Z t Z t  Z  
x(t)  c + k( )x( )d  c + k( ) c + k()x()d d
t0 t0 t0
Z t Z t Z 
=c+c k( )d + k( )d k()x()d
t0 t0 t0
" #
Z t Z t Z  Z  Z  Z 
c+c k( )d + k( )d k ( ) c + c k ( )d + k( ) k( )x( )d d
t0 t0 t0 t0 t0 t0

7
8 LINEAR DYNAMICAL SYSTEMS
" Zt Zt Z Zt Z Z #
= c 1+ k( )d + k( )d k()d + k( )d k() k( )d
t0 t0 t0 t0 t0 t0
Zt Z Z Z
+ k( )d k()d k( ) k( )x( )d
t0 t0 t0 t0
" Zt Z t 2 Z t 3 #
c 1+ k( )d + k( )d + k( )d
t0 t0 t0
Zt Z Z Z
+ k( )d k()d k( ) k( )x( )d
t0 t0 t0 t0
Keep substituting to get
R k( )d t

x(t)  ce 0 : t

Solution 3.4 By linearity,

h(t) = b0h 1 + b1 h_ 1 + b2 h1 :
Solution 3.5 Note that

x(t) = eA(t t0 ) x(t0 )

Then
       
4
=e
A 1
; 5
=e
A 1
2 1 2 2

   1  
eA = 4
2
5
2
1
1
1
2
=
3
2
1
0

   
:
0 7071 :
0 4472 2 0 :
0 7071 :
0 4472
1
=
:
0 7071 :
0 8944 0 1 :
0 7071 :
0 8944

Hence

    n   1 
x(n) = enA 1 :
0 7071 :
0 4472 2 0 :
0 7071 :
0 4472 1
0
=
:
0 7071 :
0 8944 0 1 :
0 7071 :
0 8944 0

and

     
A= :
0 7071 :
0 4472 ln 2 0 :
0 7071 :
0 4472
1
:
1 3862 :
0 6931
0 7071 : :
0 8944 0 0 :
0 7071 :
0 8944
=
:
1 3862 :
0 6931

Solution 3.6 Let

Zt
xi (t) := i ( )d
1
Then we have

x_ 1 (t) = 1( t)
x_ 2 (t) = 2 (t)
.
.
.

x_ n (t) = n (t)
y(t) = 1 (t)x1 + 2 (t)x2 +


+ n (t)xn
 9

Solution 3.7 Use PBH test: (F; G) is controllable if and only if

 
( F I G ) = A C I 0
I 0
B

has full row rank for all . Since (A; B ) is controllable, ( A I B ) is full row rank for all . Thus
( F I G ) has full row rank if and only if it has full row rank for  = 0 which implies that CA B0
has full row rank.
Solution 3.8 Obvious.
Solution 3.9 ) Suppose X is singular. We show that either (A; b) is not controllable or (c; A) is not
observable.
Since X is singular, there are u and v such that u X = 0 and Xv = 0. Hence
0

0 = u (AX + XA + bc)v = (u b)(cv):

0 0

This implies that either u b = 0 or cv = 0. Suppose cv = 0. Then

0

0 = (AX + XA + bc)v = XAv

i.e., Ker(X ) is a A-invariant subspace. Therefore there is a v 2 Ker(X ) such that cv = 0 and Av = v,
i.e., (c; A) is not observable.
Similarly if u b = 0.
0

On the other hand, suppose (c; A) is not observable. Then there is a v and a  such that cv = 0 and
Av = v.
0 = (AX + XA + bc)v = AXv + X (Av) + b(cv) = AXv + Xv = (A + I )Xv
This implies that Xv = 0 since i (A) + j (A) 6= 0. That is X is singular.
Solution 3.10 Suppose dim(w) < dim(x) dim(y), i.e., dim(w) + dim(y) < dim(x). Then
dim (span fx^g) = dim (span fPw + Qu + Ryg)  dim(w) + dim(y) < dim(x)
In other words, there is a x1 2 Rn such that x1 62 span fx^g. Since (A; B ) is controllable, there is a T > 0
and u(t) such that x(t) = x1 ; t > T . Hence it is impossible for x^(t) ! x(t) = x1 .
Solution 3.11 See page 115.
Solution 3.12 See page 66.
Solution 3.13 Note that
x_ = Ax + B1 u + B2 u_
y = C1 x + D1 u
De ne  = x B2 u. Then
_ = A + (B1 + AB2 )u
y = C1  + (D1 + C1 B2 )u
It is easy to see that (A; AB2 ) is controllable if A is nonsingular.
10 LINEAR DYNAMICAL SYSTEMS
Solution 3.14 Note that
Z n
y= a1 y( ) a~1 y(  ) + b1 u( ) + ~b1 u(  )
9
Z h >
>
i =
+ a2y(t) a~2 y(t  ) + b2 u(t) + ~b2 u(t  ) dt d
>
| {z }>
;
x2
and de ne x1 = y . Then
x_ 1 (t) = a1y(t) a~1 y(t  ) + b1 u(t) + ~b1 u(t  ) + x2 (t)
x_ 2 (t) = a2y(t) a~2 y(t  ) + b2 u(t) + ~b2 u(t  )
or
x_ 1 (t) = a1x1 (t) a~1 x1 (t  ) + x2 (t) + b1 u(t) + ~b1 u(t  )
x_ 2 (t) = a2x1 (t) a~2 x1 (t  ) + b2 u(t) + ~b2 u(t  )
y(t) = x1 (t)
Solution 3.15 (a) Suppose rank (R1 ) = r1 and rank (R2 ) = r2 . Let R1 = C1 B1 and R2 = C2 B2 with
C1 2 Rpr1 ; B1 2 Rr1 q ; C2 2 Rpr2 ; B2 2 Rr2 q ;
Then
G(s) = C1 (sIr1 + Ir1 ) 1 B1 + C2 (sIr2 + Ir2 ) 1 (sIr2 + Ir2 ) 1 B2
    
=
Ir1 B1 + Ir2 Ir2 Ir2 B2
C1 0 C2 0 Ir2 0
2 3
Ir1 0 0 B1
6 0 Ir2 Ir2 0 7
=6 7
4 0 0 Ir2 B2 5
C1 C2 0 0

(b)
       
1 1 1 1   1 0 1 0 1
G(s) = + 1 0 + +
0 s+1 1 s+1 s+2 3 10 0 0

2 3
2 3 2 0 1 0
1 1  
1 1 0 6 0 2 0 1 7
=4 1 1 5 +6 7
1 0 0 4 0 1 0 1 5
0 1
3 10 0 0
2 3
1 1 0 0 0 0
6 0 1 0 0 1 0 7
6 7
6 0 0 2 0 1 0 7
=6
6
7
0 0 0 2 0 1 7
6 7
4 1 1 0 1 0 1 5
0 1 3 10 0 0

Solution 3.16 Note that

Z  Z   
y(t) = b u
[ 1( ) ( ) a1 ()y()] + b 
[ 2( ) b  u
_ ( )] ( ) + [ _ ( )
1 1 a  a2 ( )]y( ) d d
 11

De ne x1 = y and Z 
x2 (t) = [b2 ( ) b_ 1 ( )]u( ) + [_a1 ( ) a2 ( )]y( ) d
Then
x_ 1 (t) = b1 (t)u(t) a1 (t)y(t) + x2 (t)
x_ 2 (t) = (b2 (t) b_ 1 (t))u(t) + [_a1 (t) a2 (t)]y(t)
or
   
x_ = a_ 1 (t)a1 (at)2 (t) 01 x(t) + b (tb)1 (tb)_ (t) u(t)

y(t) = 1 0 x
 2 1

Use the following relation

Z1 Z1d
Solution 3.17

AP + PA = (AeAt QeA t + eAt QeA t A)dt =
  At A t
dt e Qe dt = Q
0
 0
 0

Solution 3.18 Let Q = U U  with  = 01 0 , 1 > 0, and a unitary matrix U . Let T = U  and
partition the transformed state space realization accordingly as
 TAT  2 3
1 11 A12 B1
TB = 4 AA21 A22 B2 5
CT 1 D C1 C2 D
Then the Lyapunov equation gives
  0  A A   A A    0   C C C C 
A21 A22 + A21 A22 0 0 + C2 C1 C2 C2 = 0
1 11 12 11 12 1 1 1 1 2
0 0
which in turn gives
1A12 + C1 C2 = 0; C2 C2 = 0:
Hence C2 = 0 and A12 = 0.
For the speci c numerical problem, we have
0 0:25 0:75 0:25 1 0 0:9066 0:1072 0:4082 1
Q = @ 0:75 6:25 2:75 A ; U = @ 0:0763 0:9097 0:4082 A ;
0:25 2:75 1:25 0:4151 0:4012 0:8165
0 0:1986 0 0 1
=@ 0 7:5514 0 A
0 0 0
and 2 2:6050 7:2139 0 0:9066 0:9066 3
 TAT 1 TB  66 0:1346 0:3950 0 0:1072 0:1072 77
6
CT 1 D = 64 20::4075 2:4908 1 0:4082 2:8577 77
2625 2:2206 0 0 0 5
0:9829 1:0169 0 0 0
Solution 3.19 The Hankel singular values are listed in the following table:

1 2 3 4 5
= 2 1.572 0.641 0.209 0.061 0.017
= 4 0.992 0.689 0.418 0.245 0.156
= 20 0.589 0.549 0.497 0.449 0.416
= 100 0.517 0.510 0.500 0.500 0.483
Hence i ! 0:5 as ! 1.
12 LINEAR DYNAMICAL SYSTEMS
Solution 3.20 Use MATLAB:
 set T =??
 g = nd2sys([ 0:05; 1]; [0:05; 1]); g5 = mmult(g; g; g; g; g);
 g5 = sysbal(g5); % to avoid numerical problem.
 g10 = mmult(g5; g5); gt = nd2sys(1; [T; 1]); sys = mmult(g10; gt);
 [sysb; sig ] = sysbal (sys);

Solution 3.21 It is easy to show that

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

  1  
=
1 0
s 1
( +1)( +2)
0
s
0
0
1
0
s
1
1
:
So its McMillan form is  1 
(s+1)(s+2)
0
0 0

and its McMillan degree is 2.

Solution

3.22 Let rank R 1 = r1 and rank R2 = r2 . Then there are C1 2 R  1 ; B1 2 R 1  ; C2 2 p r r q

R p r2
; B2 2 R r2 q
such that
R1 = C1 B1 ; R2 = C2 B2
and 2 3
I 1 0 B1 r
G(s) = 4 0 2 I 2 B2 5 r
C1 C2 0
is a minimal realization. Hence the McMillan degree of G is r1 + r2 .
Solution 3.23
1
  1
  1
  1
 
G(s) = s + 1 0
0
1
3
+
s+2 0
1 0
2
+
s+3
2
1
0
0
+
s+4
0
2
0
0
2 3
1 0 0 0 0 0 1
66 0 2 0 0 0 1 0 77
66 0 0 2 0 0 0 0 77
=6 66 0 0 0 3 0 1 0 77
64 0 0 0 0 4 1 0 77
1 1 0 2 0 0 0 5
3 0 2 1 2 0 0

The McMillan degree is 5.

Solution 3.24 ??????? Find a formula for the McMillan degree of
G(s) = 1s R1 + s12 R2 ;
where Ri 2 Rpq .
Solution 3.25 Compute the transmission zeros and the corresponding zero directions of the following trans-
fer functions
 13

(a) The McMillian form is " #

1
G(s) = (s+1)(s 2)(s+3)
s2 +16s+14
s 2
p
so the zeros are 8  5 2.
(b) The McMillian form is " #
1
G(s) = (s+1)2 (s+2)
8s2 +37s+39
s+2

so the zeros are 13 8 = and 3.

(c) The only zero is 2 .
 
Note that det A zI B is equivalent to the existence of x and u such that
Solution 3.26
C D =0

     
A B x =z
I 0 x
C D u 0 0 u
which is a generalized eigenvalue problem and can be computed using a Matlab Commad: [V; D] = eig(M; N )
where    
M = CA D
B ; I
0
0
0
 
the diagonal elements of D are the invariant zeros and the columns of V are xu .
If the system is not square, then randomly augment the system to a square one and compute the zeros
for the squared system several times. Then the zeros of the original systems are the common zeros computed
from each augmentation.
Solution 3.27 Compute the transmission zeros (or invariant zeros) of the following transfer functions
(a) The only transmission zero is 0:2 with
2 3
0:0466
  66 0:0466 77
y = 6 0:1866 77
v 64 5
0:9702
0:1399

(b) No transmission zero.

(c) Two zeros are at :
2 1861 and 0:6861 with
2 3 2 3
  6 0:3015 0:8739
77 66 0:1627 77
x = 6 0:8098
u 4 0:1507 5 ; 4 0:4370 5
0:4803 0:1371

2 3 2 3
  6 0:0878 0:4636
77 66 0:1726 77
y = 6 0:4719
v 4 0:8560 5 ; 4 0:8088 5
0:1920 0:3181
14 LINEAR DYNAMICAL SYSTEMS
Solution 3.28 (a)
2 3
   A Bu 2<(z )x B 2<(z )xu
A B 2<(z )xu z 2<(z )u
=4 0 z 2<(z )u 5
C D u I C Du D
 
Apply a similarity transformation T =
I x to the above realization to get
0 1
2 3 2 3
A Ax + Bu 2<(z )x + zx B A 0 B
=4 0 z 2<(z )u 5=4 0 z 2<(z )u 5 = G(s)
C Cx + Du D C 0 D
(b) Let z1 = z2 be any pair of complex conjugate zeros of G(s), and let 1 2 C n and 1 2 C m with k1 k = 1
be the corresponding zero vectors of z1 such that
  
A z1I B 1 =0
C D 1
De ne the following parameters
=
<(z1 )1 1 ;  =  + 
2
z1 2 1 2 1

Show that
G(s) = Gc (s)Bc (s)
where  
Gc (s) = CA B 4<(z1)< [D1 (2 ) ] = k2 k
 2

2 3
<(z1 ) =(z1 ) 4<(z1 ) (=(1 ))T
4<(z1 ) (<(1 ))T
6 7
Bc (s) = 4 =(z1 ) <(z1 ) 5
=(2 )= k2 k <(2 )= k2 k
2 2
I
Chapter 4

Performance Speci cations

Solution 4.1

    2    
G 2
:= sup  G = sup max G  G
I 1 ! I ! I I
= sup max (I + G G) = sup (1 + max (G G)) = 1 + kGk21
! !
Solution 4.2
Z 1 Z 1
f (t)g(t)dt = 1 Z 1 F (j!)ej!t d!g(t)dt
1 Z1 2 1 Z 1
1
= 21 F (j!)d! ej!t g(t)dt
Z 1 1
1
= 21 F (j!)G (j!)dt:
1
Solution 4.3

Z 1 I Z =2
(j!I A) 1 d! = 1j (sI A) 1 ds + 1j rlim j j
!1 =2 (re I A) d(re )
1
1
Z =2 Z =2
= rlim
!1 (rej I A) 1 rej d = d = :
=2 =2
Note that
G (s)G(s) = F  (s) + F (s); F (s) = B  Q(sI A) 1 B
It follows that
1 Z 1 G (j!)G(j!)d! = 1 Z 1 (F  (s) + F (s)) d! = B  QB:
2 1 2 1
Solution 4.4 (a) Find a Gilbert realization for G:
2 3
1 0 0 0 1 2=3
6
6 0 3 0 0 0 5 7
7
6
G(s) = 66 0 0 2 0 10 0 7
7 = Gs + Gu
6 0 0 0 2 0 5=3 7
7
4 1 0 0 1 0 0 5
0 1 1 0 0 0
15
16 PERFORMANCE SPECIFICATIONS
Then kGs k2 = 2:211 and kGu k2 = kGu k2 = 5:069 by using the following MATLAB command
h2norm(Gs ); h2norm(cjt(Gu )):
q
Hence kGk2 = kGs k22 + kGu k22 = 5:5302. Furthermore, kGk1 = 5:46 is obtained by
linfnorm(G; 0:0001):
(b) G(s) is not stable so we can only compute its L2 norm using h2norm(cjt(G)) which gives kGk2 =
1:0801. The L1 of G is 0:7645
(c) kGk2 = 2:0616 and kGk1 = 2:972
(d) kGk2 = 1 since G is not strictly proper. kGk1 = 18:72
Solution 4.5 The steady state response is
y(t) = jG(j!)j sin(!t + \G(j!))
Then p
!max = arg max j G(j! )j = ! 1 2 2; max jG ( j! )j = p 1 :
n
! ! 2 1 2 2
Solution 4.6 Note that for a stable scalar transfer function G, the steady state response of the system with
respect to a sinusoid input u sin(!0 t + ) is ujG(j!0 )j sin (!0 t +  + \G(j!0 )). (This is the classical frequecy
response.)
The above scalar relationship can be used to the matrix case: the steady state output of yi (t) is given by
yi (t) = u1 jGi1 (j!0 )j sin (!0 t + 1 + \Gi1 (j!0 ))
+u2jGi2 (j!0 )j sin (!0 t + 2 + \Gi2 (j!0 )) +



+uq jGiq (j!0 )j sin (!0 t + q + \Giq (j!0 ))
= y1 sin (!0 t + i )
for some yi and i . yi (t) can also be written as
j (!0 t+1 ) j (!0 t+2 ) j (!0 t+q )
yi (t) = Gi1 (j!0 ) u1 e 2j + Gi2 (j!0 ) u2 e 2j +


+ Giq (j!0 ) uq e 2j
j (!0 t+1 ) j (!0 t+2 ) j (!0 t+q )
Gi1 ( j!0) u1 e 2j Gi2 ( j!0 ) u2 e 2j


Giq ( j!0) uq e 2j :
Hence
2 3 2 3
u1 ej1 u1 e j1
6 u2 ej2 7 j!0 t 6 u2 e j2 7 e j!0 t
y(t) = G(j!0 ) 664 .. 775 e 2j G( j!0 ) 664 .. 7
7
5 2j
. .
uq ejq uq e jq
Let 2 j1 3 2 j1 3
u1 e y1e
6 u e j2 7 6 y2 ej2 7
G(j!0 ) 664 .. 775 = 664 .. 775
2

. .
uq ejq yp ejp
 17

for some yi and i , and note that G( j!0 ) = G(j!0 ). Then

2 3
y1 sin(!0 t + 1 )
6 y2 sin(!0 t + 2 ) 7
y (t) = 6
6 .. 7
7
4 . 5
yp sin(!0 t + p )
Clearly, 2 3 2 3
y1 ej1 u1 ej1
6 y2 ej2 7 6 u2 ej2 7
6 7
ky^k2 = 6 .
. 7 = G(j!0 ) 6
6 .
.
7
7   (G(j!0 ))
4 . 5 4 . 5
yp ejp 2
uq ejq 2
for any 2 3 2 3
u1 u1 ej1
6 u2 7 6 u2 ej2 7
6 7 6
ku^k2 = 6 .. 7
4 . 5
= 6 .. 775  1:
4 .
uq 2 uq ejq 2
To show that the upper bound can be achieved, let G(j!0 ) have the following singular value decomposition:
G(j!0 ) = U V  = U1 V1 +

where U1 and V1 are the singular vectors corresponding to the largest singular value. Note that  (G(j!0 )) =
 (G( j!0 )) and
G( j!0 ) = U V

= U 1 V 1 +



Now write 2 3
v1 ej 1
6 v2 ej 2 7
V1 = 6
6 ..
7
7
4 . 5
vq ej q

and take ui = vi and i = i . Then

2 3
u1 ej1
6 u2 ej2 7
ky^k2 = G(j!0 ) 6
6 .. 775 = kG(j!0 )V1 k2 = k (G(j!0 )) U1 k2 =  (G(j!0 ))
4 .
uq ejq 2

Solution 4.7 Denote the impulse response of G(s) by G(t). Then yi (t) = G(t)ei .
X X XZ 1
k k =
yi 22 kG(t) k =ei 22 trace (G (t)ei ei G(t)) dt
i i i 0

Z ! Z
1 X 1
= trace G (t)

ei ei G(t) dt =

trace (G (t)G(t)) dt = kGk22
0 i 0

Solution 4.8 Invent a mechanical system consisting of masses, springs, and dashpots. Let it have at least
two force inputs and de ne at least two outputs (positions and/or velocities). Assign numerical values to
everything. Make sure it's linear (e.g., no gravity terms), time-invariant, and nite-dimensional. Find a
state model and compute the H2 -norm from the input to the output.
18 PERFORMANCE SPECIFICATIONS
Solution 4.9 Let G have the following state space realization
 
G(s) = CA D
B :

Then the Hamiltoniam matrix is

 1 

H = CA (+I +BRDR D1 DC )C BR 1 B  
(A + BR 1 D C ) ; R = I D D:
2

The following MATLAB will calculate the distance of the eigenvalues of H from the imaginary axis:
This program calls hfplot This function computes the distance:
g11=nd2sys([1,1],[1,5,6]); function [ga; dis] = hfplot(sys; gupper; gstep)
g12=nd2sys([1,0],[1,1]); [a; b; c; d] = unpck(sys);
g21=nd2sys([1,0,-2],[1,7,12]); glow=norm(d);
g22=nd2sys([1,4],[1,3,2]); [m; k] = size(d); ga = []; dis = [];
sys=sbs(abv(g11,g21), gamma = glow + gstep;
abv(g12,g22)); while gamma < gupper
gupper=5; r = gamma2  eye(k) d0  d;
gstep=0.01; ir = inv(r); ar = a + b  ir  d0  c;
[ga; dis] = hfplot(sys; gupper; gstep) H = [ar; b  ir  b0 ; c0  (eye(m) + d  ir  d0 )  c; ar0 ];
plot(ga,dis) eigH=eig(H); tmp=min(abs(real(eigH)));
dis = [dis; tmp]; ga = [ga; gamma];
gamma=gamma+gstep;
end
So kGk1  2 which is the same as from hinfnorm.
 19

0.9

0.8

0.7

0.6
distance

0.5

0.4

0.3

0.2

0.1

0
1 1.5 2 2.5 3 3.5 4 4.5 5
gamma

Figure 4.1: Distance vs.

20 PERFORMANCE SPECIFICATIONS
Solution 4.10 The results are listed in the following table. The frequency responses are calculated using log
scale, for example, w = logspace( 1; 1; 50).
 1 0.1 0.01 0.001
state space 1 3.576 35.36 353.6
50pts in [0:1; 10] 0.9852 3.5314 7.8893 8.064
100pts in [0:1; 10] 0.9852 3.565 14.458 15.723
200pts in [0:1; 10] 0.9852 3.5735 23.522 30.9735
400pts in [0:1; 10] 0.9852 3.5751 30.888 60.8893
Solution 4.11 Note that
z (s) = 1 + Pw((ss))C (s) = s (1 + P1(s)C (s))
and z (p) = 1=p. Next note that by the de nition of the Laplace transform
Z 1
z (s) = z (t)e st dt
0

Hence
Z 1 Z T Z 1
p jz (p)j jz (t)je pt dt + jz (t)je pt dt
1
= = z (t)e pt dt =
0 0 T
Z T Z 1 e PT
 e pt dt + e pt dt = +(
p
)

0 T
which gives
e PT  1

or
pT  ln 1
:
Note that p & 0 =) T % 1, thus the closer the right half plane zero to the origin, the slower the system
response.
Solution 4.12 An orthonormal basis for H2 m(s)H2 can be found by orthonormalizing
1 1 1
s + z1 ; s + z2 ; : : : ; s + zk
i.e.,  

H2 m(s)H2 = span s +1 z ; s +1 z ; : : : ; s +1 z
1 2 k
Chapter 5

Stability and Performance of

Feedback Systems
Solution 5.1 Let the output of the summation be yr ; yd and yn respectively. Then
0 1 0 0 0
10 1 0 1
@ dr A = @ A @ dr A + @ A
y H y r

y G1 0 0 y d

yn 0 G2 0 yn n

and 0 1 0 1 10 1
0
@ dr A = @ 0 A @ A
y I H r

y G1 I d

yn 0 G2 I n

So the feedback system is well-posed if

0 0 H( 1)
1
@ A
I

G1 ( 1) I 0 or I + H( 1) G2 (1) G1 ( 1)
0 G2 ( 1) I

is nonsingular and the system is internally stable i

0 0
1 1

@ 0 A
I H

G1 I =
0 G2 I

0 (I + H G2 G1 ) 1 H G2 (I+ G1 H G2 ) 1 H (I + G2 G1 H ) 1
1
@ G1 (I + H G2 G1 ) 1 (I + G1 H G2 ) 1 G1 G1 H (I+ G2 G1 H ) 1 A 2 RH
1
G2 G1 (I + H G2 G1 ) 1 G2 (I + G1 H G2 )
1 (I + G2 G1 H ) 1

If G1 and H are both stable, then the system is internally stable i

G2 (I + G2 G1 H )
1
2 RH1
Solution 5.2 Assume Xr ; Yr ; X~r ; Y~r 2 RH1 such that
Xr N + Yr M = ~ U ~ V =
+Y
I; Xr r I

Then  Yr Xr
 M 0
  0 Xr
 M U

~ ~ + ~ = I:
X r Yr 0 V X r 0 N V

21
22 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS
Solution 5.3 Any stable fraction G = MN ss ( )
( )
with proper but not strictly proper M will do. For example
s 1 s 3
N (s) = ; M (s) =
s
( + 2)( + 3) s s+3
To nd Xr and Yr such that Xr N + Yr M = I . Assume
s+ 2 s2 + 1 s + 2
Xr = 1
; Yr =
(s + 2)(s + 3) (s + 2)(s + 3)

and compare coecients to get

1 = 90 ; 2 = 180 ; 1 = 11 ; 2 = :
36

So
90 s2 + 11s 36
Xr = ; Yr =
s+3 (s + 2)(s + 3)

Solution 5.4 Let

s+
90 s2 + 11s 36 s +
Xr = ; Yr =
s+3s+ (s + 2)(s + 3) s +

Then Xr N + Yr M = I .
Solution 5.5 Note that
s s
N N + M M = 1
( 1)( + 1)
(s 2)( s s
3)( + 2)( + 3) s
p p p p p p p p
s
( s 7 12)( 7+ 12) ( + s 7 s 12)( +
7+ 12)
=
(s 2)(s 3) (s + 2)(s + 3)
p p12)(s+p7+p12)
So W (s) = (s+ 7
s
( +2)( +3) s and a normalized coprime factorization is given by
s 1 s
( + 2)( s 3)
N (s) = p p p p ; M ( s) = p p p p
( + s 7 12)( + s 7+ 12) ( + s 7 12)(s + 7 + 12)
   
Solution 5.6 This is obvious. Suppose the inverse of N X is F1 F2
M Y F3 F4 , i.e.,
  
F1 F2 N X = I
F3 F4 M Y
This gives F1 N + F2 M = I , i.e., N and M are right coprime over R[s].
Solution 5.7
2 3
1 0 0 0 1 0
6 0 2 0 0 0 1 7
6 7
6 0 0 0 0 1 0 7
G(s) = 6
6 0 0 0 2 0 1
7
7
6 7
4 1 1 0 0 0 0 5
0 0 2 1 0 0

 
F =
:
5 9675 :
0 3758 :
1 8884 :
0 1681
1:2026 4:4264 0:6588 :
0 0143
 23

2 3
4:9675 0:3758 1:8884 0:1681 1 0
6 1:2026 2:4264 0:6588 0:0143 0 1 7
6 7
6 5:9675 0:3758 1:8884 0:1681 1 0 7
  6 7
6 1:2026 4:4264 0:6588 1:9857 0 1 7
=6 7
N
6 1 1 0 0 0 0 7
M
6 7
6 0 0 2 1 0 0 7
6 7
4 5:9675 0:3758 1:8884 0:1681 1 0 5
1:2026 4:4264 0:6588 0:0143 0 1

f=lqr(a,b,c'*c,eye(2))
N=pck(a-b*f,b,c,zeros(2,2))
M=pck(a-b*f,b,-f,eye(2))
w=logspace(-1,2,300);
Nf=frsp(N,w);
Mf=frsp(M,w);
NpM=madd(mmult(cjt(Nf),Nf), mmult(cjt(Mf), Mf));
Indeed this is an identity matrix for all frequency.
Solution 5.8 The state space representations of the coprime factorizations are not unique. The results are
given for the speci c state space realizations given below for G1 and G2 .
2 3
1 0 0 0 1 1
6 0 2 0 0 1 0 7
  6 7
1 s+3 6 0 0 2 0 0 1 7
=6 7
s+1 (s+1)(s 2)
G1 (s) = 10 5 6 0 0 0 3 0 5 7
s 2 s+3 6 7
4 1 0 5=3 0 0 0 5
0 10 0 1 0 0

2 3
1:2126 3:9157 0:1815 0:3639 1 1
6 0:1577 9:4843 0:1880 0:6967 1 0 7
6 7
6 0:0548 3:5686 2:0066 0:3328 0 1 7
  6 7
6 0:2742 17:8431 0:0330 4:6641 0 5 7
=6 7
N1
6 1 0 1:6667 0 0 0 7
M1
6 7
6 0 10 0 1 0 0 7
6 7
4 0:1577 7:4843 0:1880 0:6967 1 0 5
0:0548 3:5686 0:0066 0:3328 0 1
2 3
3 2 2 1 0 0
h i 6 0 4 0 0 1 0 7
+2 6 7
G2 (s) =
s 2(s+1) 1
= 6 0 1 0 0 0 0 7
+3 s(s+4) s+1 6 7
s
4 0 0 0 1 0 1 5
1 2 2 1 0 0
2 3
3 2 2 1 0 0
6 0:2672 4:6530 1:4165 0:3375 1 0 7
6 7
  6 0 1 0 0 0 0 7
6 7
=6 7
N2
6 0:1795 0:3375 0:6139 1:2357 0 1 7
M2
6 1 2 2 1 0 0 7
6 7
4 0:2672 0:6530 1:4165 0:3375 1 0 5
0:1795 0:3375 0:6139 0:2357 0 1
24 STABILITY AND PERFORMANCE OF FEEDBACK SYSTEMS

2 3
13:6163 12:9512 0:1477 1 2 3
6 8:1085 12:4855 5:6652 3 2 1 7
6 7
6 9:1812 9:3592 3:4532 1 1 1 7
  6 7
6 1 1 1 0 0 0 7
=6 7
N3
6 2 3 4 0 0 0 7
M3
6 7
6 0:6001 3:0033 1:3638 1 0 0 7
6 7
4 1:7271 2:1197 0:4844 0 1 0 5
2:8540 1:2361 0:3950 0 0 1

2 3
2:8974 5:6922 1 2
6 1:0618 4:1854 2 1 7
  6 7
6 1 1 0 0 7
=6 7
N4
6 1 1 0 0 7
M4
6 7
4 0:0754 2:2262 1 0 5
0:9110 0:7330 0 1

Solution 5.9 Take the transpose of the transfer matrix G and perform the normalized right coprime factor-
ization. Then transpose back the coprime factors.
Chapter 6

Performance Limitations
Solution 6.1

    
 
kS k1   1 +
;  = 2 arctan !
0

Solution 6.2 We shall use the following relationship:

min (A + B ) min (A)  max (B ); min (AB )  min (A)max (B ):
1. Note that
w(s)I = w(s)(I + P (s)K (s)) 1
+ w(s)P (s)K (s)(I + P (s)K (s)) 1

and
jw(s0 )j = min (w(s0 )I )
1

= min w (s0 )(I + P (s0 )K (s0 )) + w (s0 )P (s0 )K (s0 )(I + P (s0 )K (s0 ))
1



 max w(s0 )(I + P (s0 )K (s0 )) 1 + min w(s0 )P (s0 )K (s0 )(I + P (s0 )K (s0 )) 1

1

 max w(s0 )(I + P (s0 )K (s0 )) 1 + jw(s0 )jmin (P (s0 )) max K (s0 )(I + P (s0 )K (s0 ))
 w(s)(I + P (s)K (s)) 1 1 + jw(s0 )jmin (P (s0 )) K (s)(I + P (s)K (s)) 1 1
  + jw(s0 )jmin (P (s0 )) 
2. This follows from the fact that
= = min [(I + PK )(j!)]  max (I ) + min [P (j!)K (j!)]
1

= 1 + min [P (j! )K (j! )]  1 + max [P (j! )] min [K (j! )]

i.e.,  
min [K (j!)]   1 1
:
max [P (j! )]  1

25
26 PERFORMANCE LIMITATIONS
Chapter 7

Model Reduction by Balanced

Truncation
Solution 7.1 Let T1 be the balancing transformation such that T1P T1 = ; (T1) 1 QT1 1 = . Then output
normalized realization is obtained by T = 1=2 T1 .
Solution 7.2 It is easy to show that the Hankel singular values of G are 1 = 1:6061 and 2 = 0:8561. The
H1 norm of G is kGk1 = 2:972 and the L1 norm of g(t) can be computed as
Z1
jg(t)jdt = h1 + h2 + h3 + h4 + : : :
0
R
where hi ; i = 1; 2; : : : are the variation of the step response of G shown in Figure 7.1 which gives 01 jg(t)jdt 
3:5.
2.5

2
h2

h4
1.5
step response

h3

h1
1

0.5

0
0 1 2 3 4 5 6 7 8 9 10
time

Figure 7.1: Estimating the L1 Norm of g(t)

So we have
Z1
1:6061 = 1  kGk1 = 2:972  jg(t)jdt = 3:5  2(1 + 2 ) = 4:9244:
0

Solution 7.3 It is clear that kG(s)k1 = G(0) =

P n
i=1 bi =ai and G(s) has the following state space realiza-
27
28 MODEL REDUCTION BY BALANCED TRUNCATION
tion 2 pb 3
a1 p1
6
6 a2 b2 7
7
6
G = 66 ... .. 7
7
6 p.b 7
7
a
4
pb pb


pbn 0
n 5
1 2 n
and the controllability and observability Gramians of the realization are given by
"p #
bb
P = Q = a +i aj :
i j
It is easy to see that i = i (P ) = i (Q) and
n
X n
X n
X bi = 1 G(0) = 1 kGk :
i = i (P ) = trace(P ) = 2 1
i=1i=1 i=1 ai 2
2
In particular, let ai = bi = 2i , then P = Q ! 21 In , i.e., j ! 12 as ! 1. This example also shows
that even when the Hankel singular values are extremely close they may not be regarded as repeated singular
values.
Solution 7.4 It can be veri ed that
~ e + Be Be + 12 PC
Ae P~ + PA ~ e Ce P~ =
4 N
2 3 2 3
I 0   0 0  
= 4 0 0 5 (A + A + BB  ) I0 00 0I + 4 I 0 5 (A  + A + C  C ) 00 I0 I0 = 0
0 I 0 I
so by bounded real lemma, kE11 (s)k1  2N .
Solution 7.5 A state space realization of Gd(z) is given by
 p   
Gd (z ) = (pI A) 1 (I + A) 2(I A) 1 B =: Ad Bd
2C (I A) 1 C (I A) 1 B + D Cd Dd
Then it can be veri ed that
Ad Pd Ad Pd + Bd Bd = 0; Ad QdAd Qd + Cd Cd = 0
are satis ed by Pd = P and Qd = Q where
AP + PA + BB  = 0; A Q + QA + C  C = 0:
Solution 7.6 Use MATLAB:
 T =?? and n =???
 g = nd2sys([ 0:05; 1]; [0:05; 1]); g5=mmult(g,g,g,g,g);
 g5=sysbal(g5); % to avoid numerical problem.
 g10 = mmult(g5; g5); gt = nd2sys(1; [T; 1]); sys=mmult(g10,gt);
 [sysb; sig] = sysbal(sys); sysr = strunc(sysb; n);
 err=msub(sysb,sysr); e=hinfnorm(err,0.0000001)
 29

For T = 0, all Hankel singular values are 1. Thus no reduced order model can be found.
For T = 0:01, the Hankel singular values are
i = 0:99953; 0:99803; 0:99521; 0:99047; 0:98259; 0:96907
0:94442; 0:89583; 0:79362; 0:58249; 0:22052:
Thus no reduced order model can be found.
For T = 0:1, the Hankel singular values are
i = 0:97158; 0:89465; 0:78849; 0:67228; 0:55857; 0:45297
0:35669; 0:26874; 0:18736; 0:11060; 0:03657:
A tenth order approximation gives
kG G10 k1 = 0:07313; bound = 0:07313:
For T = 1, the Hankel singular values are
i = 0:73719; 0:35146; 0:18249; 0:114322; 0:07921; 0:05744
0:04212; 0:03031; 0:02052; 0:01191; 0:00391:
A th order approximation gives
kG G5 k1 = 0:07391; bound = 0:33243:
For T = 10, the Hankel singular values are
i = 0:54386; 0:05715; 0:02054; 0:01201; 0:00814; 0:00584
0:00426; 0:00306; 0:00207; 0:00120; 0:00039:
A rst order approximation gives
kG G1 k1 = 0:08792; bound = 0:22935:
30 MODEL REDUCTION BY BALANCED TRUNCATION
Chapter 8

Hankel Norm Approximation

Solution 8.1 The inequalities i (G^ r )  i (G); i = 1; 2; : : : ; r and i (F ( s))  i+r+1 (G); i = 1; : : : ; n
r 1 hold in general. For the given example, we have
1 = 12:0256 2 = 11:34 3 = 1:1347 4 = 0:4292
r=1 1 (G^ 1 ) = 9:0920 1 (F ( s)) = 0:5089 2 (F ( s)) = 0:3564
r=2 1 (G^ 2 ) = 12:0157 2 (G^ 2 ) = 11:3145 1 (F ( s)) = 0:3039
r=3 1 (G^ 3 ) = 12:0236 2 (G^ 3 ) = 11:3377 3 (G^ 3 ) = 1:0955

31
32 HANKEL NORM APPROXIMATION
Chapter 9

Model Uncertainty and Robustness

Solution 9.1 1. max = 1= k1=(1 + Pnom )k1 = 1=15:12 = 0:0661.
2. max = 1. The exact range for stability is
> 1 or
< 11.
Solution 9.2 The following MATLAB routine can be used. (There are clearly more e ective algorithms.)
This program calls realmu This function computes real 
(M ) for a given constant
matrix M
Msys = nd2sys([1; 0:2; 1]; [1; 0:2; 11]);
function [numin; nu] = realmu(M; n)
k=200;
% M: the complex matrix
w=logspace(0,2,k);
% n: number of points computed in (0; 1]
Mf=frsp(Msys,w);
% nu: n points of sigma 2
muf = [];
% numin: the minimum of nu
for i=1:k
nu = [];
M=xtracti(Mf,i);
for i=1:n
M=var2con(M);
a=i/n;
tmp=realmu(M,400);
rm=real(M);
muf = [muf; tmp];
im=imag(M);
end
tmp = [rm; a  im; im=a; rm];
semilogx(w,muf)
s=svd(tmp);
nutmp=s(2);
nu = [nu; nutmp];
end
numin=min(nu);
When the programs are applied to the last problem, we get sup! 
(M (j!)) = 1 as shown in Figure 9.1
which is consistent with the solution obtained in the last problem.
Solution 9.3 There is no unique solution to this problem. Obviously one must choose G0 to be a minimum
33
34 MODEL UNCERTAINTY AND ROBUSTNESS

1.2

0.8

0.6

0.4

0.2

0 0 1 2
10 10 10

Figure 9.1: 
(M (j!)) vs. !

phase one in order to guarantee the stability of
. Of course, this is not necessary for robust control. In
fact, any stable G0 may be used as a nominal model and
can be unstable.
Problem 9.1 Think up an example of a physical system with the property that the number of unstable poles
changes when the system undergoes a small change, for example, when a mass is perturbed slightly or the
geometry is deformed slightly.
Solution 9.4 1. If f is a unit in RH1, then there is an  > 0 such that jf (s)j   for all Res  0. Let
g(s) 2 RH1 such that kg(s)k1 < . Then
jf (s) + g(s)j  jf (s)j jg(s)j  jf (s)j kg(s)k1   kg(s)k1 > 0; 8 Res  0:
Hence f + g is a unit in RH1 .
2. The feedback system with the plant
p = mnp +
+


n ;
n ;
m 2 RH1
p m
is internally stable i (np +
n )nk + (mp +
m )mk is a unit in RH1 . Now note that
(np +
n)nk + (mp +
m )mk = np nk + mpmk +
n nk +
mmk = f + g
with f = np nk + mp mk and g =
n nk + Deltammk . Then the system is stable if
kgk1  k
n k1 knk k1 + k
m k1 kmk k1
is suciently small which is guaranteed if k
n k1 and k
m k1 are both suciently small.
3. Give an example of a unit f (s) in RH1 such that (??) fails for the H2 -norm, that is, such that
(8 > 0) (9g 2 RH2) kgk2 <  and f + g is not a unit.
What is the signi cance of this fact concerning robust stability?
??????? we need to nd a g such that kgk1  1 and kgk2 <  for any .
Solution 9.5 The proof follows from the fact that

det I
 = det(I M
) = det(I
M )
M I
 35
Solution 9.6 The answer is no.
 
1 
1 1  2
0 1 1 3

( )=
 1 1 
1  
= s 11 11 = 1 = 4 1 0 0 5
G s 1s 1s
s s s
1 0 0
Hence a state space realization is  
_= 1 1  
y=
1
x u;
1 x
Let K = I2 , i.e., u = y . Then
_ = 2x
x

which is stable.
When  6= 0, G(s) becomes a second order system
 
2
0 0 1 13
= 64 01 0 10 00 75
1+ 1 6 7
G s ( )= 1s 1s
1 0 0 0
s s

i.e.,    
~_ = 11 10 u;
x y = 11 0 x~
Using the same controller u = y , we get

2   x~; I  2   = 2 +  
~_ =
x
1 0 1 0
and the system is not stable for any  > 0.
1 and M = s 2 be a coprime
Solution 9.7 To nd a normalized coprime factorization of G(s). Let N = s+2 s+2
factorization of G. Then
p p
 + MM = 1 1 +s 2 2 = 5+s 5 s s
s+2 s+2 s+2
s+2 s+2 s+2
NN

p5+s
Let W = s+2 . Then a normalized coprime factorization of G can be nd as G = Nn =Mn with

Nn =
N
W
= p51+ s ; Mn = W M
= ps5 +2s
 
By the robust stability condition, the system will be robustly stable for all
N
M 1 <  where
p
1= =
 
1M 1

3  s+ 5
1 (1 + GK ) =
K
n
1 1 s+1 1
p 
3 
p
= 5 1 = 5 2 = 7:0711
or  = 0:1414.
Problem 9.2 Let a unit feedback system with a controller K (s) = 1s and a nominal plant model Po(s) = s2 +s0+:21s + 5 .
Construct a smallest destabilizing
2 RH1 in the sense of k
k1 for each of the following cases:
(a) P = Po +
;
36 MODEL UNCERTAINTY AND ROBUSTNESS

(b) = Po(1 + W
) with W (s) = 0:2(s s++5010) ;

P

(c) P = N +
n , N = 2(s + 1)2 , M = s + 0:2s 2+ 5 , and
=
n
m .

2  
M +
m (s + 2) (s + 2)
Problem 9.3 Consider the system depicted here: (picture is missing)

The two controllers, K1(s) and K2(s), measure positions and generate forces. The two subsystems are
geographically separated, so there are time delays, 1 ; 2 , in communication. Assume the two time delays
are uncertain with 1  i  4. Also, assume the spring sti ness is uncertain, with 5  k  20. For slave
following and master compliance, take the performance vector to be
 
w = y1
y1
y2
w
:

Pull out the perturbations to get a system model of the form (picture is missing)

Solution 9.8 We need to check the zeros of

det (I + G(s)K ) = s + (2 + k1 + 4k2()ss ++ 1)1 +2 k1 + 4k2 2k1k2
2

So the system is stable if and only if

2 + k1 + 4k2 > 0; 1 + k1 + 4k2 2k1k2 > 0
When k1 = k2 = k, the stability conditions amount to
0:1861 < k < 2:6861:
Solution 9.9
( ) = anom(s) + (a(s) anom(s)) = 1 + a(s) anom(s)
a s
()
anom s anom (s) anom (s)

2 f1 + W (s)
(s) j k
k1  1g :
Then one can take
Ki (s) anom (s)
W (s) =
anom (s)
for any i.
 37

Stability Region
8

5 k1=k2

4
k2

−1

−2
−3 −2 −1 0 1 2 3 4 5
k1

Figure 9.2: Stability Region

Solution 9.10 1. Note that

1 1 1
y (s) = (sI + M K) M u(s) =: F (s)u(s)

Then it is easy to verify that F  (j!) + F (j!)  0.

2. We rst show that (I + F ) 1 is stable. Suppose (I + F ) 1 is not stable, then there is a 0 < k  1 such
that (I + kF ) has an imaginary axis zero, i.e., there is an x and !0 such that
(I + kF (j!0 )) x = 0

) x (I + kF (j!0 )) x + x (I + kF (j!0 )) x = 0
=

  
=) 2x x + kx [F (j!0 ) + F (j!0 )] x = 0

=) x = 0

since F is passive. Hence (I + F ) 1 must be stable.

Next to show that (I + F ) 1 (I F ) is (strictly) contractive, we note that
F +F
  2I

implies that
(I + F )(I + F )
  4I + (I F )(I F)


= ) (I + F ) 1
(I F )(I

F ) (I + F )
 1
I 4(I + F )
1 
(I + F )
1

This implies that (I + F ) 1 (I F) 1 < 1 if  > 0 and (I + F ) 1 (I F) 1  1 if  = 0.

3. Note that the following diagrams are equivalent:
38 MODEL UNCERTAINTY AND ROBUSTNESS

- 1 - K -f
- K
2
6
f
u2 6 u1
f
G  6
2  G  ?f
Then
u1 = (I + K ) 1 (I K )u2 ; u2 = (I + G) 1 (I G)u1
and (I + K ) 1 (I K ) 1  1, (I + G) 1 (I G) 1 < 1 since K is passive and G is strictly passive.
Hence by Small Gain Theorem the closed-loop system is stable.
Solution 9.11 The closed loop system is stable if and only if
 
2
2
1 + P K = (1 + P0 K ) 1 + 1KW
+ P0 K
does not have any zero in the closed right half plane. De ne
KW2
Z (s) =
1 + P0 K
and
ZI = =Z (j!); ZR = <Z (j!):
Then 1 + P K has a right half zero if and only if there is an admissible
such that
1 + Z (j!)
2 (j!) = 0
for some !. Or, equivalently
ZR =
2 + ZI <
2 = 0
and
1 + ZR <
2 ZI =
2 = 0
Solving the above equations to get
<
2 = jZZRj2 ; =
2 = jZZIj2
Thus the system is stable if and only if for any given frequency ! either
ZR
jZ j2 >
or
ZI
jZ j2 >
Write in another form:  
ZR ZI
max jZ j2 ; jZ j2 > 1:
 39
P 
Problem 9.4 Let P = P21 PP1222 2 RH1 be a two by two transfer matrix. Find sucient (and neces-
11

sary if possible) conditions in each case so that Fu (P;
) is stable for all possible stable
that satis es the
following conditions, respectively
1. at each frequency
<
(j!)  0; j
(j!)j <
2. at each frequency
<
(j!)ej  0; j
(j!)j <
where   0.
3. at each frequency
<
(j!)  0; =
(j!)  0; <
(j!) + =
(j!) <
Problem 9.5 Let G(s) 2 RH1 be a m  m symmetric transfer matrix, i.e., GT (s) = G(s), and let
(s) 2
RH1 be a diagonal perturbation, i.e.,
0  ( s) 1
1
B

(s) = B
 (s) 2 CC :
@ ... A
m
Show (I G
) 1 2 RH1 for all k
k1  if and only if kGk1 < 1= . (Hint: Note that for a com-
plex symmetric matrix M = M T 2 C mm , there is a unitary matrix U and a diagonal matrix  =
diag(1 ; 2 ; : : : ; m)  0 such that M = U U T (see Horn and Johnson, page 204). For detailed discus-
sion of this problem, see Qiu (1995). )
Solution 9.12 Since P and P have the same number of unstable poles, the controller K stabilizing P will
0 0
stabilize P if we can show that det(I + PK ) and det(I + P K ) have the same number of encirclements around
0
the origin. (This is the Nyquist stability criterion.) Now
det(I + PK ) = det I +
WP K (I + P K )
det(I + P K )
0 0
1
0

and
WP K (I + P K ) < 1, so det(I + PK ) and det(I + P K ) have the same number of encirclements.
0 0
1
0

Problem 9.6 Give appropriate generalizations of the last problem to other types of uncertainties.
Solution 9.13 A state space realization of P is 0
2 1 0 0 1 03
66 0 1 0 1 1 77
P = 66 0 0 3 0 2 77
4 1 0 1 0 05
0

0 1 0 0 0
1. max = 1= K (I + P K ) 1 = 1=1:152 = 0:8681.
0
1

2. Use the Program in Problem 9.2 to get max = 1=1:1399 = 0:8773.

3. ???????
Problem 9.7 Repeat the last problem with
" s 1 5 +1 s #
P0 = s
( +1)2
1
s
( +1)2
s
1 :
s
( +1)2 s
( +1)2
40 MODEL UNCERTAINTY AND ROBUSTNESS
Chapter 10

Linear Fractional Transformation

 
Solution 10.1 1. M =( I 0) and N = I

I
.

 0  0 0
2. M = I

0 I
and N = I

0 0 I
.

0 0 0 01 0 0 01 I

= @0 0A @ 0 0 00 CA
=B
I
I
3. M I I and N .
0 0 0 I
0 0
I

0 0 0 01 0 0 01 I

= @0 0A =B
@ 0 0 00 CA
I
I
4. M I I and N .
0 0 0 I
0 0
I

0 0 01
0 0 0 0 0 01
I

BB 0 0C
0C
I

= @0 A =B BB 00 C
I

0 0 0C
I
5. M I

0 0 0 0
I I

0
and N

@0 C .

0A
I
I
I

0 0 I

Solution 10.2

G = +
(
I P I P
) 1 = Fu(
) M;

with
 
M = P

P
I

I
:

Solution 10.3 Note that

   [1 +
( )] ( ) [1 +
( )] ( )     
z1

z2
= 11 11 12
s 12 g s

[1 +
21 ( )] 21( ) [1 +
22( )] 22 ( )

s g s
s

s
g

g
s

s
w1

w2
= g11

g21
g12

g22
w1

w2

2
32 0 3
 0 
0 66
11

12 77 66 0 77 
I

+ 0 g11 g12

0 g21 g22 4
21 54 0 5

I
I w1

w2

22 0 I

41
42 LINEAR FRACTIONAL TRANSFORMATION
Solution 10.4 Let 2 3
A B1 B2 
A^ B^

P (s) = C1 D11 D12 ; K (s) = C^ D^
4 5
C2 D21 D22
Find state space realizations for F`(P; K ) and F` (P; D^ ). (Hint: Use the formula on page 267 and page 268.)
Solution 10.5
2   3

A  
0 B2  
B1 
6
M^ (s) = 64 C1 0 D12 D11 7
7
0
 0 0
 I 5
C2 I D22 D21 + E
2 3
A B2 D211 C2 B1 D211 B2 B1 D211 D22
F`(M;
^ E 1
)=4 C1 D11 D211 C2 D11 D211 D12 D11 D211 D22 5
D211C2 D211 D211 D22
Chapter 11

Structured Singular Value

Solution 11.1 (a) Not true. For example,
   
0

M = 0 0 ;
= 0
1 0
2
Then 
(M ) = 0 but M 6= 0.
(b) Not true. For example,
     
M1 = 00 0 ; M2 = 0 00 ;
=
01
0
2
Then 
(M1 ) = 
(M2 ) = 0 but 
(M1 + M2 ) 6= 0.
(c) True. Follows from the de nition.
(d) 
(I ) = 1 since (I ) = (I ) = 1.
   

= I2 , M = 10 101 , and N = 101 01 . Then 
(MN ) =
(e) Not necessarily true. Let

 101 10   102 > (M )
(N )  10
1.
10 1
(f) Not necessarily true. As above.
Solution 11.2 Follows from
  
det I M 11 M12
1 0  = det(I M11
11 ) det(I M22
2)
M21 M22 0
2
Solution 11.3

M = rand(7; 7); blk = [3; 0; 2; 2; 2; 0];

[bnds; rowd; sens; rowp] = mu(M; blk); pert = unwrapp(rowp; blk)
Solution 11.4 The solution to this problem was shown in Freudenberg [2] and Nett et al [1]. Note that for
two block uncertainty problem
   

(M ) = d> dI 0
inf0  0 I M21 0 0 M 12 I=d 0
 
0 I 
 
= d> 0
inf0  M21=d 0 dM 12 0
= inf  I 0 M21=d 0I 0 dM 12
  d>0
= d> inf0  dM0 12 M210 =d = d> inf0 fd(M12); (M21 )=dg
p
= (M12)(M21):
43
44 STRUCTURED SINGULAR VALUE
Solution 11.5 Note that for two block  problem we have
 dI 0  M I=d 0
 

(M ) = inf  11 M12
d>0 0 I M21 M22 0 I
 M11 dM12

= d>
inf0  M21 =d M22
 M11 0
  0 dM 
= d>
inf0  12
0 M22 + M21 =d 0
Hence  M   
 11 0 0
inf  dM12  
(M )
0 M22 d>0
M21 =d 0
 M11 0
  0 dM 
 inf0  M =d 0 12
0 M22 + d> 21
Solution 11.6 Let
be all diagonal full blocks and M be partitioned as M = [Mij ] where Mij are matrices
with suitable dimensions. Show that

(M )   ([kMij k]) (=  ([kMij k]))
where (
) denotes the Perron eigenvalue.
Solution 11.7
0 J1 0 0 0 1
inf DMD 1 p = infnn
B0
DB
J2 0 0 C 1
C D = (M )
D2C nn D2C @ 0 0 .. 0 A
.
0 0 0 Jm p
where Ji are the Jordan blocks. It is now easy to show that
infki Ki Di Ji Di 1 p = (Ji )
Di 2C
and thus the result follows.
Solution 11.8 The solution of this problem is given in J. Stoer and witzgall (1962), F. L . Bauer (1963),
Safonov (1982), and Safonov and Doyle (1984).
Solution 11.9 Let x and y be partitioned compatibly with the
:
0 x1 1 0 y1 1
x=B
B x2 CC ; y = BB y2 CC
@ .. A . @ .. . A
xm+F ym+F
Then

(M ) = max (MU ) = max (y  Ux)
U 2U U 2U
= max u y x +


+ um ym  xm + y U1 xm+1 +


+ y UF xm+F
U 2U 1 1 1 m+1 m+F
= max ju y x j +


+ jum ym xm j + ym +1 U1 xm+1 +


+ ym +F UF xm+F
U 2U 1 1 1
= jy1x1 j +


+ jym xm j + kym+1 k kxm+1 k +


+ kym+F k kxm+F k
 45

where the maximizing ui and Ui are given by


ui = e j \yi xi ; Ui = ky ym+ki xkmx +i k + U~i
m+i m+i
such that U~i xm+i = 0, ym +i U~i = 0, and Ui is unitary. It is also easy to show that max
2B
(M
) is
achieved by

i = e j\y x ;
i = ky ym+ki xkmx +i k :
i i

m+i m+i
Solution 11.10 Note that for i = 0; 1; : : :; k

kx1 k2 + kz0 k2  2 (kx0 k2 + kd0 k2 )

kx2 k2 + kz1 k2  2 (kx1 k2 + kd1 k2 )
..
.
kxk+1 k2 + kzk k2  2 (kxk k2 + kdk k2 )

Adding the above inequalities together to get

kX+1 Xk X k Xk
kxi k2 + kzi k2  2 ( kxi k2 + kdi k2 )
i=1 i=0 i=0 i=0
or
k k k
kxk+1 k2 kx0 k2 + (1 2 ) X kxi k2 + X kzi k2  2 X kdi k2
i=0 i=0 i=0
Since the above inequality is true for any k, it is concluded that x 2 `2 and z 2 `2 and furthermore
limk!1 kxk k = 0. Hence we have
1 1 1
kx0 k2 + (1 2 ) X kxi k2 + X kzi k2  2 X kdi k2
i=0 i=0 i=0
or
kz k22 + (1 2 ) kxk2  2 kdk2 + kx0 k2
2 2
Let a discrete time system be given by
    
xk+1 = A B xk
zk C C dk
The interpretation of the above is that the state space realization satis es
 
A B  < 1:
C D
Solution 11.11 Clearly, K needs to stabilize P0 .
(a) By internally stability condition, the system is stable if and only if for all
i :
1 + P (j!)K (j!) 6= 0; 8!
m
1 + W1 T
1 + W2 SK
2 6= 0
*
k jW1 T j + jW2 KS j k1  1
46 STRUCTURED SINGULAR VALUE
To show that 1+W1T
1 +W2 SK
2 6= 0 implies k jW1 T j + jW2 KS j k1  1. Suppose k jW1 T j + jW2 KS j k1 >
1, i.e.,
:= jW1(j!0 )T (j!0)j + jW2 (j!0)K (j!0)S (j!0)j > 1
for some frequency !0 . Let  0 and  0 be such that
j!0
\ + j! = \W1(j!0)T (j!0)
0
j!0
\ + j!
0
= \W2(j!0)K (j!0)S (j!0)
and let

1 = 1 + ss ;
2 = 1 + ss :
Then k
i k1 < 1 and
j \W1 (j!0 )T (j!0 )

1 (j!0) = jW (j! )T (j!e )j + jW (j! )K (j! )S (j! )j
1 0 0 2 0 0 0
j \W ( j! )K (j! )S (j! )

2 (j!0) = jW1 (j!0)T (ej!0)j + jW2 (j!0)K (j!0)S (j!0)j
2 0 0 0

and
1 + W1 (j!0)T (j!0)
1 (j!0) + W2(j!0 )K (j!0)S (j!0 )
2(j!) = 0
i.e., j!0 is a closed-loop pole. Therefore, k jW1 T j + jW2 KS j k1  1 is necessary for the closed-loop
stability.
(b)
W3 W3 S
Tzd =
1 + P K = 1 + W1 T
1 + W2 KS
2 :
kTzdk1  1
*
WS
1+W1 T
1 +3 W2 KS
2  1; 8!;
i
*
jW3 S j  j1 + W1 T
1 + W2 KS
2j ; 8!;
i
*
jW3 S j  1 jW1 T j jW2 KS j ; 8!;
*
jW3 S j + jW1 T j + jW2 KS j  1; 8!;
*
kjW3 S j + jW1 T j + jW2 KS jk1  1.
On the other hand, suppose kjW1 T j + jW2 KS jk1  1 (which is necessary for robust performance) but
kjW3 S j + jW1 T j + jW2 KS jk1 > 1
i.e., there is an !0 such that
0 := jW3 (j!0 )S (j!0 )j + jW1 (j!0 )T (j!0 )j + jW2 (j!0 )K (j!0 )S (j!0 )j > 1:
Then there exists a > 1 such that
jW3 (j!0 )S (j!0 )j + 1 jW1 (j!0 )T (j!0 )j + 1 jW2 (j!0 )K (j!0 )S (j!0 )j > 1:
It is also easy to construct stable
1 and
2 such that k
i k1 = 1= < 1 and

1(j!0) = e \W1 (j!0 )T (j!0 ) = ;
2 (j!0) = e \W2 (j!0 )K (j!0 )S(j!0 ) = :
Thus
 47
jW3 (j!0 )S (j!0 )j > 1 1 jW1 (j!0 )T (j!0)j 1 jW2 (j!0 )K (j!0 )S (j!0 )j
+
jW3 (j!0 )S (j!0 )j > j1 + W1 (j!0 )T (j!0 )
1 (j!0 ) + W2 (j!0 )K (j!0 )S (j!0 )
2 (j!0 )j
+
jW3 (j!0 )S (j!0 )j
j1 + W1 (j!0 )T (j!0)
1 (j!0 ) + W2 (j!0 )K (j!0 )S (j!0 )
2 (j!0 )j > 1
+
kTzw k1 > 1.
This is a contradiction.
Solution 11.12 Draw the block diagram as below. Then it is easy to show that
0z 1 0 10 1 0 1
@ z12 A = @
W1 T W1 T W1 T W3 d1 d1
W2 KS W2 KS W2 KSW3 A @ d2 A = M @ d2 A
z S S SW3 d d
with  W1 T W1 T
  W1 T

M11 = = ( 1 1)
W2 KS W2 KS W2 KS
Hence by problem ??, 
s (M11 ) = jW1 T j + jW2 KS j. Next note that

 Ds M11 Ds 1

=   dW1 T  ( 1=d 1)

W2 KS
s  dW T   dW T 
= max 1 ( 1=d 1) ( 1=d 1) 1
W2 KS W2 KS
s  dW T  dW T 
p 1 1
= 1 + 1=d2 max
W2 KS W2 KS
s  dW T   dW T 
p 1 1
= 1 + 1=d2 max
W2 KS W2 KS
p q
= 1 + 1=d2 jdW1 T j2 + jW2 KS j2
q
= jdW1 T j2 + jW2 KS j2 + jW1T j2 + jW2 KS=dj2
The minimum is achieved by
d2 =
jW2 KS j
jW1 T j
and
= jW T j + jW KS j
inf  Ds M11 Ds 1 1 2
Ds
Similarly, it is easy to see that 0 1
W1 T
M =@ W2 KS A ( 1 1 W3 )
S
and

2
 DMD 1 = jW1 T j2 + jW2 KS j2 + jW3 S j2 + jW2 T d1 =d2 j2
+jW1 KSd2=d1j2 + jW3 T d1j2 + jW2 S=d2j2 + jW3KSd2j2 + jW1S=d1j2
48 STRUCTURED SINGULAR VALUE
It can be shown, which is not obvious, that the minimizing di are given by
jW1 S j jW2 S j
d21 = ; d22 =
jW3 T j jW3 KS j

and
inf 1

2 = jW1 T j2 + jW2 KS j2 + jW3 S j2
D  DMD
+2jW1 T j jW2 KS j + 2jW1 T j jW3 S j + 2jW2 KS j jW3 S j
= (jW1 T j + jW2 KS j + jW3 S j)2

- W2 z2-
2 d2 d
?
W3
- W1 z1-
1
-e - K - P0 - ?ed2- ?e - ?e -z
6

Figure 11.1: Diagram for Problem????

Solution 11.13 Note that

v
u F F
uX X
(DMD 1 ) =  ([xi yj di =dj ]) = t kxi k2 kyj k2 d2 =d2
i j
i=1 j=1
With the given di , it is easy to see that
v v
u F F u F F
uX X uX X
(DMD 1 ) = t kxi k2 kyj k2 d2 =d2
i j= t kxi k kyj k kxj k kyi k
i=1 j=1 i=1 j=1
v
u F F F
uX X X
= t kxi k kyi k kxj k kyj k = kxi k kyi k
i=1 j=1 i=1
which is 
(M ) by Problem ??.
Solution 11.14 9.11. Clearly, K needs to stabilize P0 . By internally stability condition, the system is stable
if and only if for all
i :
1 + P (j!)K (j!) 6= 0; 8!
m
1 + W2 SK
2 6= 0
m
(< f
2 g + j = f
2 g) + 1=(W2 SK ) 6= 0
m
< f
2 g + < f(1=(W2 SK ))g + j (= f
2 g + = f(1=(W2 SK ))g) 6= 0
m
< f(1=(W2 SK ))g > ; = f(1=(W2 SK ))g >
 49

Solution 11.15 Consider M 2 C 2n2n to be given. Let
2 be a n  n block structure, and suppose that
2 (M22 ) < 1. Suppose also that Fl (M;
2 ) is invertible for all
2 2 B
2 .
For each > 1, nd a matrix W and a block structure
^ such that
(Fl (M;
2 )) <

2max
2B
2
if and only if

^ (W ) < 1;
where  is the condition number.
50 STRUCTURED SINGULAR VALUE
Chapter 12

Parameterization of Stabilizing
Controllers
Solution 12.1 Note that K stabilizes Pi is equivalent to Pi stabilizes K . It is well known that all stabilizing
controllers for a plant K can be parameterized as an LFT
P~ = Fu (J; Q)
for some J and some Q 2 H  Pi stabilizes K , there must be a Qi such that Pi = Fu (J; Qi ). Let
 1 . Since
I
= maxfkQi k1 g and P = I
J . Then

P~ = F (J; Q) = F (P; Q= ):
u u

De ne
P = fF (P;
) : k
k1  1g
u

Then it is clear that fPi g  P .

Solution 12.2 (Internal Model Control (IMC)) Suppose a plant P is stable. Then it is known that all
stabilizing controllers can be parameterized as K (s) = Q(I P Q) 1 for all stable Q. In practice, the exact
plant model is not known, only a nominal model P0 is known. Hence the controller can be implemented as
in the following diagram.
r -e -e - Q - P -y
6 6
P0 

The control diagram can be redrawn as follows. This control implementation is known as Internal Model
Control (IMC). Note that no signal is feeded back if the model is exact. Generalize the IMC to unstable
systems.

r -e - Q - P -y
6
- P0 - ?e

51
52 PARAMETERIZATION OF STABILIZING CONTROLLERS
Chapter 13

Riccati Equations
Solution 13.1 1. Note that     
B 0 U U ^
C A V = V B
where B^ is similar to B . Since U is nonsingular, we have
    
B 0 I = V UI ^
U BU 1
C A VU 1 1

or     
B 0 I I ^
X = X U BU
1
C A
 
Multiply X I from the left to get
  
X I BC 0A XI = 0 =) XB + C AX = 0:
 

On the other hand, suppose X is a solution to the Sylvester equation, then

        
B 0 I = B = B = I B
C A X C AX XB X
Write B^ = U 1 BU for the Jordan form of B and set
V := XU
Then     
B 0 I I
C A X U = X BU
or     
B 0 U U
V = V U BU
1
C A
 
i.e., the columns of U V are the eigenvectors of M associated with the eigenvalues of B . Since U is
nonsingular, X = V U 1 .
2. The necesssity is easy. Suppose X is a solution and let
 
T = IXn I0m
Then    
B 0 T =T B 0
C A 0 A
53
54 RICCATI EQUATIONS
Solution 13.2
Z t 
A P (t) + P (t)A = A eA  QeA + eA  QeA A d
0

Z t
d heA QeA i  = eAt QeAt Q = P_ (t) Q
=
0 d

_ t;  ) = (t;  )A(t).
Solution 13.3 Easy by the fact that (
Solution 13.4 Note that  
_ t) = HeHt =
( A R 
Q A
or

_ _ _ _
       
11 12 = A R ; 21 22 = Q A 
and
P (11 + 12 P0 ) = 21 + 22 P0
Then

P_ (11 + 12 P0 ) = P (_ 11 + _ 12 P0 ) + _ 21 + _ 22 P0

   
  I
= P A R  P0 + Q A  P0

  I
   
= P A R 11 + 12 P0 +  Q A  11 + 12P0
 
21 + 22 P0 21 + 22P0
So    
P_ = P A R
  I + Q A
 I 
P P = PA PRP Q A P
Solution 13.5 Note that
 
_ t) = HeH t
(T ) = I; ( ( T) = A R (t)
Q A
or
_ = A + R ; _ = Q
11 11 12 21 11 A 21
Note also that
P 11 = 21
Hence
P_ 11 = _ 21 + P _ 11 = Q11 + A 21 + P (A11 + R12 )
i.e.,
P_ = Q + A P + PA + PRP
Solution 13.6 J -spectral factorization and H1 control???
Chapter 14

H2 Optimal Control
Problem 14.1 The following block diagram shows a plant p , actuators a , sensors s , and controller :
G G G K

u
- Ga - Gp
v
-
r
- K

6 Gs 
Assume Gp (s), Ga (s), and Gs (s) are strictly proper and K (s) is required to be proper. The signals r(t),
u(t), and v(t) are 2 dimensional. In addition to internal stability, the performance requirements, roughly,
are that u(t) should not be too large and r(t) v(t) should be small. Each component of r(t) is given by
8
>< 10t; 0 t < 1
10; 1 t < 2
r i( ) =
t
>: 10t + 30; 2 t < 3
0; 3t :

Set this problem up as an H2 optimization problem. Take state models as required and regularize as necessary.

55
56 H2 OPTIMAL CONTROL
Chapter 15

Linear Quadratic Optimization

57
58 LINEAR QUADRATIC OPTIMIZATION
Chapter 16

H1 Control
Solution 16.1 Bilateral hybrid telerobot. The setup is shown in Figure 16.1. Two robots, a master,

?
f
-j -
h
m
v m - s v

s  j
6 6e
G G
f
K

f m s
f

Figure 16.1: Bilateral hybrid telerobot.

G m , and a slave, Gs , are controlled by one controller, K . A human provides a force command, fh , to the
master, while the environment applies a force, fe , to the slave. The controller measures the two velocities,
vm and vs , together with fe via a force sensor. In turn it provides two force commands, fm and fs , to the

master and slave. Ideally, we want motion following (vs = vm ), a desired master compliance (vm a desired
function of fh ), and force re ection (fm = fe ).
For simplicity of computation we shall take Gm and Gs to be SISO with transfer functions
1 1
^
g m (s) = s
; s (s) =
^
g
10s
:

We shall design K for two test inputs, namely, fe (t) is the nite-width pulse

e (t) =
f
10; 0  02
t :
(16.1)
0; t > 0:2;

indicating an abrupt encounter between the slave and a sti environment, and fh(t) is the triangular pulse
8
< 2t; 0  1
t

h( ) = :
f t 2t + 4; 1  2
t (16.2)
0; t > 2;

to mimic a ramp-up, ramp-down command.

The generalized error vector is taken to have four components: the velocity error vm vs ; the compliance
error fh vm (for simplicity, the desired compliance is assumed to be vm = fh); the force-re ection error
fm fe ; and the slave actuator force. The last component is included as part of regularization, that is, to

penalize excessive force applied to the slave. Introducing four weights to be decided later, we arrive at the
59
60 H1 CONTROL

generalized error vector 2 3

wv (vm vs )
z = 64
6 wc (fh vm ) 77 :
wf (fm fe ) 5
ws fs
The Laplace transforms of fe and fh are not rational:
f^e (s) = 10 f^h(s) = s22


2
1 e 0:2s ; 1 e s :
s
To get a tractable problem, we shall use second- and third-order Pade approximations,

e Ts  1
Ts + (Ts)2  1 + Ts + (Ts)2 
2 12 2 12
and   
e Ts  1
Ts + (Ts)2 (Ts)3
1+
Ts + (Ts)2 + (Ts)3  :
2 10 120 2 10 120
Using the third-order approximation for f^e (s) and the second-order one for f^h(s), we get
   
0:2 0:23 s2 0:2s (0:2s)2 (0:2s)3
f^e (s)  20
2
+
120
1+
2
+
10
+
120
=: g^e (s)
,
s s 2 2
f^h(s)  2 1+ +
2 12
=: g^h (s):
Incorporating these two pre lters into the preceding block diagram leads to Figure 16.2. The two exogenous

wh- vm- ?
Gh -j - Gm vs Gs  j Ge we
6 K 6
fm fs
Figure 16.2: Telerobot with pre lters.

inputs wh and we are unit impulses. The vector of exogenous inputs is therefore
 
w = wwhe :
The control system is shown in Figure 16.3, where z and w are as above and
2 3
vm  
y = 4 vs 5 ; u = ffms :
fe
Beginning with state models for Gh ; Gm ; Gs ; Ge , namely,
       
Ah Bh ; Am Bm ; As Bs ; A e Be ;
Ch 0 Cm 0 Cs 0 Ce 0
 61

 z  w
G

y u
- K

Figure 16.3: Telerobot con gured in the standard form.

with corresponding states xh ; xm ; xs ; xe , using the interconnections in Figure 16.2, and de ning the state
2 3
xm
6 xs 7
x = 64 xe 7
5
xh
lead to the following state model for G:
2 3
Am 0 0 Bm Ch 0 0 Bm 0
6
6
0 As Bs Ce 0 0 0 0 Bs 7
7
6
6
0 0 Ae 0 0 Be 0 0 7
7
6
6
0 0 0 Ah Bh 0 0 0 7
7
2
A B1 B2
3 6
6
Cm Cs 0 0 0 0 0 0 7
7
4 C1 0 D12 5 :=
6
6
Cm 0 0 Ch 0 0 0 0 7
7 : (16.3)
C2 0 0
6
6
0 0 Ce 0 0 0 I 0 7
7
6
6
0 0 0 0 0 0 0 I 7
7
6 7
6 7
6
6 0 0 Ce 0 0 0 0 0 7
7
4 0 Cs0 0 0 0 0 0 5
Cm 0 0 0 0 0 0 0
For the data at hand, D21 = 0, so (A2) fails. Evidently, the condition D21 = 0 re ects the fact that no
sensor noise was modelled, that is, perfect measurements of vm ; vs ; fe were assumed. Let us add sensor noises,
say of magnitude . Then w is augmented to a 5-vector and the state matrices of G change appropriately so
that the realization becomes 2 3
A 0 B1 B2
4 C1 0 0 D12 5 :
C2 I 0 0
Some trial-and-error is required to get suitable values for the weights:
wv ; wc ; wf ; ws ; :
The MATLAB functions h2syn and h2lqg can be used to compute the optimal controller. Play with the
weights and try to get reasonable performance. Include plots in your solution.
Solution 16.2 The generalized plant G is given by
2 3 2 3 2 3
W 0 WP W   0 0 0
4 0 0 1 5 = 4 0 5 1 0 P + 4 0 0 1 5
F 2 FP F 0 2 0
62 H1 CONTROL

 W = nd2sys([1]; [2:5=pi; 1]);

 W = mmult(W; W );
 sys1 = abv(W; 0; F );
 sys2 = sbs(1; 0; mmult( 1; P ));
 D = [0; 0; 0; 0; 0; 0:01; 0; 0:01; 0];
 G = madd(mmult(sys1; sys2); D);
 [K; Tzw; gamma] = hinfsyn(G; 1; 1; 0; 10; 0:0001)

 w = logspace( 1; 3; 400);
 Tzwf = frsp(Tzw; w);
 vplot(0 liv; lm0; Tzwf )
 Tf = vnorm(Tzwf );
 vplot(0 liv; lm0; Tf )
 axis([0:1; 1000; 0:00001; 1])
 grid
2 3
:
1 2566 :
1 2566 0 0 0 0 0 0
6
6 0 :
1 2566 0 :
1 1193 :
1 5841 :
1 1210 0 0 7
7
6
6 0 0 :
6 2832 :
2 5029 :
3 5423 :
2 5066 0 0 7
7
Gs
6
( )=6
0 0 0 0 01: 0:0282 0 0 :
0 9985 7
7
6
6 0 0 0 :
0 0282 20 0 0 :
1 4132 77
6
6 :
1 1210 0 0 0 0 0 0 0 7
7
4 0 0 0 0 0 0 0 0 01: 5
0 0 :
2 5066 0 0 0 :
0 01 0

We get : ,
opt = 0 08

:4598  106 (s + 0:01)(s + 20)(s + 6:2832)(s + 3:9612)

K (s) = (s +810578)( s + 619:0752)(s + 35:381)(s + 1:3224)(s + 1:1862)
 (799 :7664(s + 0:01)(s + 20)(s + 6:2832)(s + 3:9612)
s + 619:0752)(s + 35:381)(s + 1:3224)(s + 1:1862)
The closed loop frequency responses are shown in Figure 16.4.
Solution 16.3 Note that K robustly stabilizes the uncertain system if and only if K stabilizes P0 and
WK (I P0 K ) 1 1  1= :
Without loss of generality, assume that W (s) is stable and minimum phase. Then K stabilizes P0 if and
only if WK stabilizes P0 W 1 . Let
P0 W 1 = Ps + Pu
such that Ps is stable and Pu is antistable. Let K1 = WK (I Ps WK ) 1 then
WK (I P0 K ) 1 = WK (I Ps WK Pu WK ) 1
  1
= WK (I Ps WK ) 1
I Pu WK (I Ps WK ) 1
 63

0
10

−1
10 the largest singular value of Tzw

T21
−2
10

−3
10

−4
10 T22 T12 T11

−5
10 −1 0 1 2 3
10 10 10 10 10

Figure 16.4: Frequency responses of the closed-loop

K1 (I Pu K1 ) 1 :
=

Clearly, K stabilizes
 P0 if and only if K1 stabilizes Pu .
A B
Let Pu = C 0 be a controllable and observable realization (so A is stable). Let

Pu = NM 1
= ~ M 1 N~
then there exist U;
~ V~ ; U; V 2 RH
1 such that
V~ M UN
~ = I; MV
~ NU
~ = I:

All stabilizing controller can be parameterized as

K1 = (U + MQ) 1 (U + NQ) 1 ; Q 2 RH1
and
K1 (I Pu K1 ) 1
=( U + MQ)M~
Let X > 0 and Y > 0 be the stabilizing solutions to
XA + A X XBB  X = 0; Y A + AY Y C  CY = 0
and let F = B X and L = Y C  . Then
2 3
 
M =4 F A + BF B
 

A + LC L B


N I 5; M
~ N
~ =
C I 0
C 0

 
U= A + BF L
F 0
64 H1 CONTROL

and M  M = I; M~ M~  = I . Hence

K WK (I P0 K )
inf 1
1 = inf
K1 K1 (I inf 1 (U + MQ)M~ 1
1 = Q2RH Pu K 1 ) 1

= Q2RH inf 1 k(M U + Q)k1 = k(M U )kH

inf 1 k(U + MQ)k1 = Q2RH
It is easy to show that  
(M  U ) = A XB
L 0
Moreover, X and Y are the controllability and observability Gramians of (M  U ) . Therefore, k(M  U ) kH =
(XY ). On the other hand, it is easy to see that X 1 is the controllability Gramian of Pu and Y 1 is the
observability Gramian of Pu . Hence (XY ) = 1=min(X 1 Y 1 ) = 1=min(Pu ). Therefore we have
 min (Pu ) = min ([P0 W 1 ]+ ) = 1=(XY ):
We can also use the standard H1 control setup. We have
2 3
 A 0 B 
G(s) = I0 PIu = 4 0 0 I 5
C I 0

0 W
with controller K1 or G(s) = I P with controller K .
0
Solution 16.4 ??????????????? Let P = P0 (I + W
) with k
k1 < and assume P and P0 have the
same number of right half plane poles. Find the largest such that the closed-loop can be robustly stabilized.
Again by small gain theorem, a controller K robustly stabilizes the uncertain system if and only if K
stabilizes P0 and
(I KP0) 1KP0W 1  1=
We have  
G(s) = P0 W PI0 0

Solution 16.5 Let G 2 RH1 and let G = NM 1 be a coprime factorization with M inner, i.e., M M = I .
Then
inf kG Qk1 = Q2RH
Q2RH1
inf 1 NM 1 Q1
  
= Q2RH
inf 1 kN QM k1 = Q2RH
inf 1 F` N I ;Q
M 0 1
Solution 16.6 This problem was discussed in detail in the paper by Aoki, Ushida, and Kimura (1995). The
relationship between the given for computing the central controller and the actually H1 norm ( 1 ) is
shown in Figure 16.5. Obviously, 1 is not monotonic w.r.t. .
Solution 16.7 The problem can be set in a LFT from:
2 3 2 3
 
z1 =

M 1 0 M 1  w1  M  A BF B
z2 P I P
4 w2 5 ; N = 4 F I 5
w3 C 0
2 3

M 1 0 M 1  A B 0 B
G(s) = P I P =4 F I 0 I 5
C 0 I 0
 65

H−infinity Norm of the Central Controller Is NOT Monotonic w.r.t. gamma

1.23

1.22

1.21
H−infinity norm with central controller

1.2

1.19

1.18

1.17

1.16

1.15

1.14
1 1.2 1.4 1.6 1.8 2 2.2
gamma

Figure 16.5: A counter example to monotonicity properties of H1 central controllers.

The Matlab function lqr is used to determine that

 
F = 1 339:7965 :0001 :0087 :
The function pck is then used to create G(s) and this system is entered as a parameter for the function
[k; t; r] = hinf syn(G(s); 1; 1; 0; 5; :000001) :
This function determines that the optimal is 3.5913 and that the optmal controller has poles at :004613 
j 1:5385, :01436 and 1:6371
107 , and zeros at -.0012, :00462  j 1:539.
Solution 16.8
    
z1 W2 K (I + P K ) 1
W2 K (I + P K ) 1
w1
z2
= W1 (I + P K ) 1 W1 ( I + P K ) 1 w2

2 3 2 32 3
z1 W1 W1 W1 P w1
4 z2 5 = 4 0 0 W2 5 4 w2 5
z3 I I P w3
2    3

6
50 s=1:245+1
s=0:007+1
50 s=1:245+1
s=0:007+1
50 s=1:245+1
s=0:007+1
0:5(1 s)
(s+2)(s+0:5)
7
0 G(s) = 6
0:1256
4
7
50 s=0:502+1
s=2+1

1 1 0:5(1 s)
(s+2)(s+0:5)
2 3
:5 0 0 0 0 1
6
6 0 2 0 0 0 1 77
(Gilbert Realization) = 66 :0212 :1065 0 :007 :348 :348 :178 7
6
7
6 1 :281 :281 0 77
4 0 :7496 0 0 0 :5004 5
0:5 1 0 1 1 0
As in the previous problem, this G(s) is used as a parameter for hinf syn and the optimal is found to be
1.418, the poles of the optimal controller are -0.007, -2.7246 and 6:5438
104 and the zeros are -0.5, -2.
66 H1 CONTROL
 
For the second situation where
=


2 , the Matlab function dkit is used to nd the optimal
1

controller. The values of  with respect to frequency are shown in the the gure below.

CLOSED−LOOP MU: CONTROLLER #5

1.4

1.2

1
MU

0.8

0.6

0.4

0.2 −1 0 1 2 3 4 5 6
10 10 10 10 10 10 10 10
FREQUENCY (rad/s)

Figure 16.6: Frequency response of the closed-loop

The data after ve iterations is displayed in the following table:
Iteration Summary
Iteration # 1 2 3 4 5
Controller Order 3 13 13 13 13
Total D Scale Order 0 10 10 10 10
Achieved 1:418 1:298 1:298 1::298 1:298
Peak  V alue 1:416 1:298 1:297 1:298 1:298
The order of D was chosen to be 5, resulting in a 13th order controller. Controllers of higher order do not
signi cantly improve performance. The nal value for is 1.298 and the poles and zeros for the controller
after 5 iterations are listed below:
0:007 3:4503
:494 483:2229
1:3321 609:2501
Poles : 1:8709 923:0608
2:8353 1949:7079
2684:8153 9913:8806
10230:871
:4939 564:343
:5 1087:346
Zeros : 1:336 1910:324
1:932 6439:409
2 9898:695
3:02 538787:6
Chapter 17

H1 Control: General Case

Problem 17.1 Let G(s) 2 H1 be a square transfer matrix and > 0. Show that G is (extended) strictly
positive real if and only if ( I G)( I + G) 1
1 < 1.
Problem 17.2 Let G(s) 2 H1 be a square transfer matrix and y = Gu. Then G is (extended) strictly
positive real if and only if
Z 1  Z 1
ky + uk2 ky uk2 dt   ky + uk2 dt
0 0

for some  > 0. (Note that ky + uk2 ky uk2 = (y u + u y). One implication of the result is
that a strictly positve real problem can be converted to an equivalent strictly bounded real problem. The
transformation induced from this result can also be applied to nonlinear systems.)
Solution 17.1 Note that
" p # !
( I X )( I + X ) 1
= F` pI 2I ;X
2 I I

Then nding a controller K such that F`(G; K ) is (extended) strictly postive real is equivalent to nding a
controller such that " p # !
F` pI2 2 I ; F (G; K )
`
I I

is strictly bounded real. Using the star product formula on page 268, we have
" p # !
F` pI2 I ; F (G; K )
2
` = F` (G;
~ K)
I I

where 2 p 3
A pB1 R 1 C1 = B R 1=
2 1 B2 pB1 R 1 D12 =
G~ (s) = 4 2R 1 C1 (I
p D11 = )1R
1 2R 1 D12 5
C2 D21 R 1 C1 = 2D21 R = D22 D21 R 1 D12 =
and R = I + D11 = . It is noted that
 p    
I B1 =( p2 )A B1 Rp 1 C1 = j!I B2 pB1 R 1 D12 = A j!I B2
0 R= 2 2R 1 C1 2R 1 D12
=
C1 D12
 p    
A B1 R 1 C1 = j!I p 2B1 R 1 = Ip 0
p A j!I B1
C2 D21 R 1 C1 = 2D21 R 1 = C1 = 2 R 1 = 2 =
C2 D21
67
68 H1 CONTROL: GENERAL CASE

Hence the standard assumptions for the H1 control are satis ed if

   
A j!I B2 ; A j!I B1
C1 D12 C2 D21
are both of full rank.
Solution 17.2 Assume D11 = 0 and D12 = 0 in the realization of G. Then
Q + (I + sN )F` (G; K ) = F`(G;
~ K)

with 2 3
A B1 B2
G~ = 4 C1 + NC1 A Q + NC1 B1 NC1 B2 5
C2 D21 D22
Chapter 18

H1 Loop Shaping
Solution 18.1 See problem ???? in Chapter 9. Note that a normalized coprime factorization for G is given
by    
Nn 1 1
= s + p5 s 2
Mn
and it is easy to compute that  
Nn
Mn
= 0:2298
H
Hence max = 0:9732.
Solution 18.2 ??????? In Corollary 18.2, nd the parameterization of all H1 controllers.

Solution 18.3 It is straightforward to verify that

  
K
(I + P K ) 1M~ 1 = U + MQ  :
I V + NQ
Next note that   ~ ~  ~ ~  
~ =I ) =I
M U V U V U M U
N V ~
N M ~
N ~
M N V
which implies
~
MV ~ = I:
NU
Furthermore, if P = N M 1 = M~ 1N~ are normalized coprime factorizations, we have
          
N N N N
~ M~ = N~ M~ ~ M~ = I :
M M M M
N~ M~ N N

Hence           
M N
1
bP;K = U
+ M
Q = ~ ~
U
+ M
Q
V N
1  N

M V N
1
= R +Q :
I
1
Solution 18.4 Note that
K I( + P K ) 1 1  1=bP;K , K stabilizes P~ = P +
a; k
a k < bP;K
P K (I + P K )
1 ~
1  1=bP;K , K stabilizes P = (I +
m )P; k
m k < bP;K
KP (I + KP )
1 ~
1  1=bP;K , K stabilizes P = P (I +
m ); k
m k < bP;K
(I + KP ) 1 1  1=bP;K , K stabilizes P~ = P (I +
f ) 1; k
f k < bP;K
(I + P K ) 1 1  1=bP;K , K stabilizes P~ = (I +
f ) 1P; k
f k < bP;K
69
70 H1 LOOP SHAPING

Solution 18.5 Think P as the controller and K as the plant.

Problem 18.1 Let an uncertain plant be given by
P = s +s 2+s + 1 ; 2 [1; 3];  2 [0:2; 0:4]
2

and let a nominal model be

P = s2 +s 2+ s0+ 1
0

1. The largest possible k
add k1 ; = 4:2686 and k
mul k1 = 1:1415

 
2. The largest possible
n
m 1 = 0:4552.
3. in part (2), let (N ; M ) be a normalized coprime factorization of P . Find the largest possible

n
m  1 .
P = N=M with
s+2
N = s2 + 1:9576 ; M = s2 + 0:6s + 1
s + 2:2361 s2 + 1:9576s + 2:2361
is a normalized coprime factorization.
 
4. Find the optimal nominal 0 and 0 such that the largest possible k
addk1 , k
mul k1 , and
n
m 1
are minimized respectively.
Problem 18.2 Let P = s s . Design (a) a precompensator W of order no greater than 2 such that the
(
10
1)
crossover frequency !c  2 and bopt (WP ) is as large as possible; (b) nd the optimal loop shaping controller
K = K1 W .
Problem 18.3 Let P = s s s . Design (a) a precompensator W of order no greater than 2 such that the
100(1
( +10)
)

crossover frequency !c  1 and bopt (WP ) is as large as possible; (b) nd the optimal loop shaping controller
K = K1 W .
Solution 18.6 Note that the Riccati equation can be written as
X (A + BF ) + (A + BF ) X + C  C + F  F = 0
 
i.e., X is the observability gramian of N
M . Let P be the controllability gramian, i.e.
P (A + BF ) + (A + BF )P + BB  = 0
and without loss of generality, assume X and P are balanced with


X=P = 
 :
1
2

 
Then partition the realization of N
M accordingly
2 3
2
A + BF B
3 A11 + B1 F1 A12 + B1 F2 B1
6 A21 + B2 F1 A22 + B2 F2 B2 7
4 C 0 5 = 64 C1 C2 0
7
5
F I F1 F2 I
Then F1 = B1 1 and
 (A + B F ) + (A + B F )  + C C + F F = 0
1 11 1 1 11 1 1 1 1 1 1 1
 71
This implies the r-th order balanced reduced order model
2 3
N^ = 4 A11 +C B1 F1 B01
 

M^ 1 5
F1 I
is also a normalized coprime factorization.
Solution 18.7 (Reduced Order Controllers by Controller Model Reduction, McFarlane and Glover
 
A B
(1990), Zhou and Chen (1995)) Let G(s) = C 0 = M~ 1 N~ be a normalized left coprime factorization
and let K (s) be a suboptimal controller given in Corollary 18.2 (with performance ). Let K = UV 1 be a
right coprime factorization 2 3
 
U = 4 A BBC X1 YIC 5
 

V B  X1 0
^ V^ 2 RH1 be approximations of U and V . De ne
and U;

U
 
U^
 := V V^ 1
and Kr = U^ V^ 1 . Show that Kr is a stabilizing controller for G if  < 1 and
   
Kr (I + GK ) 1 M~ 1 = Kr (I + GK ) 1  I G  < :
I I 1 1 
r r
1
Solution 18.8 (a) Note that K stabilizes the family:
n o
F := (M~ +
~ M ) 1 (N~ +
~ N ) :
~ M
~ N 1 < 
 

and n o
(M~ r +
^ M ) 1 (N~r +
^ N ) : 
^ M
^ N  1 <    F :
Hence K stabilizes M~ r 1 N~r with stability margin of at least  :
 
K (I + G K ) 1 M~ 1  ( ) 1 :
r
I r
1

~ M
~ N := M~ M~ r N~ N~r . Then
   
(b) Denote
W 1 W  = (N~ +
~ N )(N~ +
~ N ) + (M~ +
~ M )(M~ +
~ M )

  
~ 
= I +  N~ M~ 

N + 
N
M  N~  M
M
 

+
N
M 
N

M

and W 1  1 +  and kW k1  .
1 1
1 

(c) This follows from the fact that Gr = (W M~ r )(W N~r ) 1 is a normalized left coprime factorization and
1
   
inf K1 (I + G K ) 1 (W M~ ) 1  K (I + G K ) 1 (W M~ ) 1
K1 I r 1 r
I r
1
r
1
 
K (I + G K ) 1 M~ 1 W 1 W 1 1
I r r 1  
1
and    
K1 (I + G K ) 1 M~ 1  K1 (I + G K ) 1 (W M~ ) 1
kW k1 :
I r 1 r
1 I r 1 r
1
72 H1 LOOP SHAPING

(d) Note that  

K1 (I + G K ) 1 M~ 1   1 kW k
r 1 1
I r
1
rn

implies that K1 stabilizes

 
F1 := (M~ r +
M ) 1 (N~r +
N ) :


M
N  
1 < kW k1 :
rn

Then the result follows from the fact that

 
(M~ +
~ M ) 1(N~ +
~ N ) : 

~ M
~ N
 
1 < kW k1   F1
rn

i.e., K1 stabilizes M~ 1 N~ with stability margin of at least kWrnk1 :

K1 (I + GK ) 1 M~ 1  rn  1 :
   

I 1 kW k1
1
(e) This is obvious because
   
inf K2 (I + Gr K2 ) 1 M~ 1  K (I + G K ) 1 M~ 1  ( ) 1
K2 I r
1 I r r
1
(f) Similar to (d).
 
Solution 18.9 Let G(s) = CA B0 = M~ 1N~ be a normalized left coprime factorization and let K (s) be
a suboptimal controller given in Corollary 18.2 (with performance ):
  
K (s) = A BB BX1X Y C C Y C
1 0
where   1
2 2
X1 = 2 1 Q I 2 1Y Q
and
AY + Y A Y C  CY + BB  = 0
Q(A Y C  C ) + (A Y C  C ) Q + C  C = 0:
Suppose Y and Q are balanced,i.e.,
Y = Q = diag(1 ; : : : ; r ; r+1 ; : : : ; n ) = diag(1 ; 2 )
and let G(s) be partitioned accordingly as
2 3
A11 A12 B1
G(s) = 4 A21 A22 B2 5 :
C1 C2 0
  1
Denote Y1 = 1 and X1 = 1 1 I 1 1
2
2 2
2 2 . Show that
   
Kr (s) = A11 B1 BB1 XX1 Y1 C1 C1 Y1 C1
1 1 0
 73

is exactly the reduced order controller obtained from the last problem with balanced model reduction procedure.
(It is also interesting to note that
Q = X (I + Y X ) 1

where X = X   0 is the stabilizing solution to

XA + A X XBB  X + C  C = 0:
Hence balancing Y and Q is equivalent to balancing X and Y . This is called `Riccati Balancing', see
Jonckheere and Silverman, 1983.)
Solution 18.10
 Apply the
 controller reduction methods in the last three problems respectively to a satellite
system G(s) = A B where
C 0
2 3 2 3
0 1 0 0 0

; B = 64 1:73190 10
6 0 7 6 5 7
A = 64 0
0
0
0
0
0
1
7
5
7
5 ;
0 0 :
1 5392 2 0:003  1:539 3:7859  10 4

 
C = 1 0 1 0 ; D = 0:
Compare the results.
74 H1 LOOP SHAPING
Chapter 19

Controller Order Reduction

Problem 19.1 Apply the frequency weighted balanced model reduction procedure to
U^
   

~Nn M~ n  U V^ V 1
where N~n ; M~ n; U , and V are given on page 503. Show that the weighted controllability Gramian P and
observability Gramian Q can be obtained from
(Ak + Bk Fk )P + P (Ak + Bk Fk ) + Bk Bk = 0
   
 Q 1 Q 12 Q 1 Q
Ao Q Q + Q Q Ao + C0 Co = 0
12
12 12
where 
~ 1
Ao := A + BB2 DRk R1 C C2 A + BB2RR 1CDk C
1


k 2 k k 22 k
 
Co := R C2 R D22 Ck Fk
1 1

Suppose K = UV 1 is an observer-based controller, i.e.,

Ak := A + B2 F + LC2 + LD22 F; Fk := (C2 + D22 F ); Lk := (B2 + LD22)
Bk := L; Ck := F; Dk := 0:
Then show that P and Q can be obtained from
(A + B2F )P + P (A + B2F ) + LL = 0
(A + LC2) Q + Q(A + LC2) + C2 C2 = 0
Problem 19.2 Let G be a generalized plant and K be a stabilizing controller.Let
 = diag(

 p ;
k ) be a
suitably dimensioned perturbation and let Tz^w^ be the transfer matrix from w^ = w to z^ = zz in the
w
1 1
following diagram

z  w
G 
zw z
1 - - f w 1
W 1 K

- (K^ K )W

75
76 CONTROLLER ORDER REDUCTION
Let W; W 1 2 H1 be a given transfer matrix. Show that the following statements are equivalent
  
I 0
1. 
0W 1 Tz^w^ < 1;
2. kF` (G; K )k1 < 1 and W 1 Fu (Tz^w^ ;
p ) 1 < 1 for all  (
p )  1;
  
3. W 1 Tz1w1 1 < 1 and F` I0 W0 1 Tz^w^ ;
k < 1 for all (
k )  1;
1
Problem
 19.3 In ^ 1
 the part3 of Problem 19.2, if we let
k = (K K )W , then Tz1 w1 = G22 (I KG22 ) and
F` I0 W0 1 Tz^w^ ;
k = F` (G; K^ ). Thus K^ will stabilizes the system and satis es F`(G; K^ ) 1 < 1
if k
k k1 = (K^ K )W  1 and the part 2 of Problem 19.2 is satis ed by a controller K . Hence to
1
reduce the order of the controller K , it is sucient to solve a frequency weighted model reduction problem if
W can be calculated. In the single input and single output case, a \smallest" weighting function W (s) can
be calculated using the part 2 of Problem 19.2 as follows
jW (j!)j  sup jFu (Tz^w^ (j!);
p )j:
(
)1
P

Repeated Problem ?? and Problem ?? using the above method. (Hint: W can be computed frequency by
frequency using  software and then tted by a stable and minimum phase transfer function.)
Problem 19.4 One way to generalize the method in Problem 19.3 to MIMO case is to take a diagonal W
W = diag(W1 ; W2 ; : : : ; Wm )
and let W^ i be computed from
jW^ i (j!)j  sup jeTi Fu (Tz^w^ (j!);
p )j
(
P )1
where ei is the i-th unit vector. Next let (s) be computed from
j (j!)j  sup jW^ 1 Fu (Tz^w^ (j!);
p )j
(
P )1
where W^ = diag(W^ 1 ; W^ 2 ; : : : ; W^ m). Then a suitable W can be taken as
W= W ^:
Apply this method to Problem ??.
Problem 19.5 Generalize the procedures in Problem 19.3 and Problem 19.4 to problems with additional
structured uncertainty cases. (A more general case can be found in Yang and Packard [1995]).
Chapter 20

Structure Fixed Controller

77
78 STRUCTURE FIXED CONTROLLER
Chapter 21

Discrete Time Control

Solution 21.1 Using B = (A I ) 1 (A + I ), we get
B  P~ + PB
~ + Q = 0; P = 21 (B I ) P~ (B I )
On the other hand, using C = (A I )(A + I ) 1 , we get
C  P + PB + 2(A + I ) T Q(A + I ) 1
:
=0

Problem 21.1 Let a discrete time transfer function be given by

 
G(z ) = A B
C D
and assume D = 0. Derive a computational procedure to compute the H1 norm of the transfer matrix G(z )
without the assumption that A is nonsingular.

79
80 DISCRETE TIME CONTROL
Bibliography
[1] C. N. Nett and J. A. Uthgenannt, \An explicit formula and an optimal weight for the 2-block structured
singular value interaction measure," Automatica, vol 24, No. 2, pp. 261-265, 1988.
[2] J. S. Freudenberg, \The general structured singular problem with two blocks of uncertainty." Proc.
1986 Allerton Conf., Monticello, Illinois, October 1986.
[3] N. Aoki, S. Ushida, and H. Kimura, \On -characteristic of H1 control systems," Proc. 34th IEEE
Conf. Dec. Contr., New Orleans, Louisiana, December 1995, pp. 2562-2567.
[4] Jonckheere, E. and L. M. Silverman (1983). \A new set of invariants for linear systems - applications
to reduced order compensator design," IEEE Transactions on Automatic Control, Vol 28, No. 10, pp.
953-964.
[5] Qiu, L. (1995). \On the robustness of symmetric systems," Proc. of 34th IEEE Conf. Dec. Contr.,
New Orleans, Louisiana, pp. 2659.
[6] Qiu, L., B. Bernhardsson, A. Rantzer, E. J. Davison, P. M. Young, and J. C. Doyle (1995). \A formula
for computation of the real stability Radius," Automatica, Vo. 31, No. 6, pp. 879-890.
[7] K. Zhou and J. Chen \Performance bounds for coprime factor controller reductions," Systems and
Control Letters, Vol. 26, No. 2, pp. 119-127, 1995.

81