Introduction to MVC

Definition---Properness and strictly properness

A system G(s) is proper if all its elements {g ij ( s)} are proper, and strictly proper

if all its elements are strictly proper.

Definition---Causal
A system G(s) is causal if all its elements are causal, and not causal if all its
elements are noncausal.
Definition---Poles
The eigen values λi , i=1,---,n of the system G(s) are called the pole of the system.
n
The pole polynomial π ( s ) is defined as: π ( s) = ∏ ( s − λi ) , where π ( s ) is the
i =1

least common denominator of all non-identical-zero minors of all order of G(s).

Example:

 ( s − 1)( s + 2) 0 ( s − 1)2 
1  
G ( s) =  
( s + 1)( s + 2)( s − 1)
 −( s + 1)( s + 2) ( s − 1)( s + 1) ( s − 1)( s + 1) 
 
The minor of order 2:
1 2 −( s − 1)
1,2
G1,2 = 1,2
; G1,3 = 1,2
; G2,3 =
( s + 1)( s + 2) ( s + 1)( s + 2) ( s + 1)( s + 2)2
The least common denominator of the minors:
π ( s) = ( s + 1)( s + 2)2 ( s − 1)
Definition---Zeros
If the rank of G(z) is less than the normal rank, z is a zero of the system. The
zero polynomial Z(s) is the greatest common divisor of the numerators of all
order-r minors of G(s), where r is the norminal rank of G(s) provided that these
minors have all been adjusted in such a way as to have the pole polynomial
π ( s ) as their denominator.
( s − 1)( s + 2) 2( s − 1)( s + 2) −( s − 1) 2
1,2
G1,2 (s) = 1,2
; G1,3 (s) = 1,2
; G2,3 (s) =
π ( s) π (s) π ( s)
So, Z ( s ) = ( s − 1)
According to this definition, the zero polynomial of a square G(s) is:
det{G ( s )} = 0
Notice that, in a MIMO system, there may be no inverse response to indicate the

1
 presence of RHP-zero. For example, in the MIMO system of the following:

1  1 1
G ( s) = ; G (0.5) = 0; σ {G (0.5)} = 0
(0.2 s + 1)( s + 1) 1 + 2 s 2 


1  0.45 0.89  1.92 0   0.71 −0.71

H

G (0.5) =
1.65 0.89 −0.45  0 0   0.71 0.71 
144244 3 1424 3 14 4244 3
U Σ VH

Transfer functions for Closed loop MIMO systems

1. Cascade rule. For the cascade interconnection of G1 ( s ) and G2 ( s) in the
following figure(a), the overall transfer function matrix is G=G1G2
2. Feedback rule. With reference to the positive feedback system in Figure (b):

v = ( I − L) −1 u = ( I − G2G1 ) −1 u

3. Push-through rule. G1 ( I − G2G1 ) −1 = ( I − G1G2 ) −1 G1

G1 − G1G2G1 = G1 ( I − G2G1 ) = ( I − G1G2 )G1

⇒ ( I − G1G2 )−1 G1 ( I − G2G1 ) = G1
⇒ ( I − G1G2 )−1 G1 = G1 ( I − G2G1 ) −1

2
4. MIMO rule.
(1). Start from the output, write down the blocks as moving backward by the most
direct path toward the input
(2). When exit from a feedback loop, include a term (I-L)-1 for positive feedback
(or, (I+L)-1 for negative feedback. Notice that L is the evaluated against signal
flow starting at the point of exit from the loop.
Example: Consider the following block diagram:

z =  p11 + p12 K (1 − P22 K ) −1 p21  w

5. Consider the closed-loop system:

The following relationships are useful:

( I + L) −1 + L( I + L)−1 = S + T = 1
G ( I + KG ) −1 = ( I + GK ) −1 G
GK ( I + GK ) −1 = G ( I + KG )−1 K = ( I + GK ) −1 GK
T = L( I + L)−1 = ( I + L)−1 L = ( I + L−1 )−1

3
Singular Values and Matrix Norms

1. Vector Norms
A real valued function • defined on a vector space X is said to be a norm on

X, if it satisfies the following properties:

(1) x ≥ 0 ;

(2) x = 0 only if x = 0 ;

(3) α x = α x , for any scalar α ;

(4) x + y ≤ x + y ;

for any x ∈ X and y ∈ X .

 n 1/ p 
The vector p-norm of x ∈ X is defined as: x p
=  ∑ xi  , for 1 ≤ p ≤ ∞. In
 i =1 
particular p = 1, 2, ∞ we have:

n n
x 1 = ∑ xi ; x2= ∑ (| x |) = max xi .
2
i ; x ∞ 1≤i ≤ n
i =1 i =1

H o&& lder inequality:

1 1
xT y ≤ x p
y q, + =1
p q

Notice that:

(1) x2≤ x1≤ n x 2

;

(2) x ∞
≤ x 2 ≤ n x ∞;

(3) x ∞
≤ x1≤ n x ∞

4
2. Matrix norms:
A matrix norm is defined as a function valued that satisfies the following:

(1) A ≥ 0 for all A ∈ R m×n with equality only if A=0;

(2) α A = α A for all α ∈ R, A ∈ R m×n ;

(3) A + B ≤ A + B for all A, B ∈ R m×n

Let A = [aij ] ∈ R m×n , the vector induced p-norms is defined as:

Ax
A p
= sup p

x≠0 x p

In particular:
m
A 1 = max ∑ aij (column sum) ;
1≤ j ≤ n
i =1

A 2 = λmax ( A* A ) ;
n
A ∞
= max ∑ aij (row sum)
1≤ j ≤ n
j =1

Another popular matrix norm is the Frobenius norm, i.e.:

m n
A F
= ∑∑ | a
i =1 j =1
ij |2

The p-norms have the following important property for every A ∈ R m×n :

(1) AX p
≤ A p
x p ; p=1,2, ∞ ;

(2) A2≤ A F
≤ n A 2
;

(3) max
1≤i ≤ m ;1≤ j ≤ n
aij ≤ A 2 ≤ mn { max
1≤i ≤ m ;1≤ j ≤ n
aij };
1
(4) A ∞
≤ A2≤ m A∞ ;
n
1
(5) A1≤ A2≤ n A1
m

5
Lemma 1: Let x ∈ F n and y ∈ F m .

(1) Suppose n ≥ m . x = y iff ∃U ∈ F n×m s.t. x = Uy, and U *U = I ;

(2) Suppose n = m . x* y ≤ x y . Moreover, equality holds iff x = α y

for some α ∈ F or y = 0 ;

(3) x ≤ y iff ∃∆ ∈ F n×m with ∆ ≤ 1 s.t. x = ∆y . Furthermore,

x ≤ y iff ∆ < 1 ;

(4) Ux = x for any appropriate dimensioned unitary matrix U .

Lemma 2: Let A and B be any appropriate dimensioned matrices.

(1) ρ ( A) ≤ A p , ( p = 1, 2, ∞, F ) ;

−1
(2) AB ≤ A B ; A −1 ≥ A ;

(3) UAV = A , and UAV F

= A F , for any unitary matrices U and V ;

(4) AB F
≤ A B F , and AB F
≤ B AF.

Lemma 3: Let A be a block partitioned matrix with:

 A11 A12 A1q 

A A21 A2 q 
A= = A 
21

   ij 
 
 Am1 Am 2 Amq 

Then, for any induced matrix p-norm,

 A11 A12 A1q 

 p p p

 A A22 A2 q 
Ap≤  p
21 p p

 
 
 Am1 Am 2 Amq 
 p p p
p

Further, the equality holds if the F-norm is used.

6
7
Singular Value Decomposition

Consider a fixed frequency ω where G ( jω ) is a l × m complex matrix.

Denote G ( jω ) as G for simplicity. Any matrix G may be decomposed into

G = U ΣV H .
Where,
Σ is an l × m matrix with k = min{l , m} non-negative singular values, σ i .
arranged in descending order along its main diagonal; the other entries are
zero. The singular values are the positive square roots of the eigenvalues of
G H G , and G H is the conjugate transpose of G. That is:

σ i (G ) = λi (G H G )

y = G⋅d ;
Let u be one of the eigenvector of G.

{ }
2
2 2 2
d = vi ⇒ vi =α2 vi ,1 + vi ,2 + L vi ,n
2

y = G ( vi ) = Gvi = λi vi
⇒ y = α 2λi2 vi
2 2

y λi vi
∴ 2
= = λi
d 2
vi

On the other hand, it can be shown that the extreme values of Gd 2 / d 2

are λi1/ 2 G H G  , which is known as the singular value of G, in the

directions of eigenvaectors of G H G .

y y y
Since, σ = min 2
≤ 2
≤ max 2
=σ
d 2
d 2
d 2

As a result, σ ≤ λi ≤ σ

The singular values can be considered as the extreme gains of the MIMO

8
 system, which are local maximal values with respect to the direction of
inputs. For example, consider the gain matrix at a specific frequency:
5 4 
G= 
3 2 
The gains with respect to the direction of d are given in the following
figure:

Typical singular values with respect to frequency are as shown in the

following figure:

Theorem 1: Let A ∈ F m×n . There exist unitary matrices:

U = [u1 , u2 , L , um ] ∈ F m×m ; V = [v1 , v2 , L , vn ] ∈ F n×n

Σ 0 
such that: A = U ΣV * , Σ =  1 
 0 0
where,

σ 1 0 L 0
0 σ L 0 
Σ1 =  ; σ 1 ≥ σ 2 ≥ L ≥ σ p ≥ 0 ; p = min {m, n}
2

M M O M 
 
 0 0 L σ p 

9
[Proof]

Let σ = A , and assume m ≥ n .

Then, from the definition of A , i.e.:

Az
A p
= sup p
; ( p = 1, 2, ∞ ) ⇒ Az p
= A p
z p
for some z
z p

In other words, there exists a vector z ∈ F n such that

Az = σ z = σ z

By the Lemma

( x = y iff there is a matrix U ∈ F m×n such that x = Uy and U *U = I )

there is a matrix U% ∈ F m×n such that U% *U% = I and

Az = U% (σ z ) = σ Uz
% .

z %
Uz
Let: x = ∈ F n and y = ∈Fm
z %
Uz

Az σ Uz
% σ 2Uz
% σ 2Uz
% %
Uz
We have: Ax = = = = = =y
z  Az  Az σ Uz
% %
Uz
 
 σ 

Let U = [ y, U1 ] ∈ F m×m ; V = [ x, V1 ] ∈ F n×n be unitary.

Thus,
A1 = U * AV = [ y, U1 ]*[ Ax, AV1 ]
 y * Ax y* AV1   σ y* y y * AV1 
= * = * 
σ U1 y U1 AV1 
* *
U
 1 Ax U 1 AV1

σ w* 
= 
0 B 
Since,
2
1 
0
A1   = σ 2 + w* w A1 ≥ σ 2 + w* w
2
⇒
M 
 
0 2

10
 and σ = A = A1 , we conclude that w = 0 .

An obvious induction argument gives: U * AV = Σ . The σ i is the i-th

eigen value of A, and ui and v j are i-th left singular vector and j-th right

singular vectors, respectively. It is obvious to see:

Avi = σ i ui and A*ui = σ i vi

or, A* Avi = σ i2 vi and AA*ui = σ i2ui

Hence, σ i2 is an eigen value of AA* or A* A , ui is an eigen vector of

AA* , and is an eigen vector of A* A . The following notations for

singular values are often used:
σ ( A) = σ max ( A) = σ 1 = max Ax
x =1

σ ( A) = σ min ( A) = σ p = min Ax
x =1

Lemma 1: Suppose A and D are square matrices.

(1) σ ( A + ∆) − σ ( A) ≤ σ ( A);

(2) σ ( A∆) ≥ σ ( A)σ ( ∆);

1
(3) σ ( A−1 ) = if A is invertible;
σ ( A)

11
12
Lemma 3:

(1) σ ( A) ≤ λ ≤ σ ( A)

(2) λ ≤ A p

σ ( A)
(3) Let κ ( A) = = condition number of A , and Ax = b . If
σ ( A)

13
 δx  δb 
A( x + δ x) = (b + δ b) , then: = κ ( A) ×  
x  b
 

(4) Let A% → A + δ A and x% = x + δ x , where A and x satify Ax = b .

Then:

δx κ ( A)
≤
x δA
1 − κ ( A)
A

Thus, when κ ( A) is large and matrix A is almost singular, a very

small change of A will make the possible range of δ x large.

14
Applications of SVD

15
16
17
18
Problem of small singular values:
Very small singular values in a multivariable system are analogus to very
small gains in a conventional siso system. It requires very large
controller gains and results in excessively large controller actions. The
typical presence of constraints in the manipulated variable and noise in
the sensor makes it difficult even for siso system. In the context of mimo
system, the additional complications presented by hidden loops,
interactions make the problem even more severe.
A general rule of thumb to measure the small singular value is the
magnitude of the noise in the signal. Singular values are equal or less
than the magnitude of sensor noise should be assumed degenerate.

Problem of large singular values:

Large singular values are not as serious a problem as small values. It
requires very small controller gains. This results in small controller
outputs, which can be easily become lost in the resolution of the
manipulator. The symptomatic behaviors are cyclic responses, which
never settling down to a reasonable steady state.
A general rule of thumb concerning large singular values is that all
singular values that are equal to or greater than the reciprocal of the
valve resolution should be avoided.

Determining Good Sensor Locations

Consider, for example, the ethonal-water distillation column as shown in
Figure 3.Assume that the first level objective is to control two column
temperatures by manipulating D and Q. The basic concern is to
determine which combination out of 1225 possible combinations of
sensor locations from the point of view of column control.

19
The 50 × 2 gain matrix is as shown in figure 4.

20
21
From the SVD result in Figure 5 and is plotted on Figure 6. The largest
elements in U1 and U2 occurs on stage 18 and 13. On the other hand, if
we use abs(U1)-abs(U2) as criterion, as shown on Figure 7, the largest
differences suggests that stage 18 and stage 13 are good choices.

22
23
Selection of Proper Manipulated variables

For example, in the design of control of a control for a distillation

column, four manipulated variables are typically be considered:
D distillate flow rate
L reflux flow rate
B bottoms flow rate
Q steam rate to the reboiler
Two out of the four have to control levels (i.e. accumulator and column
base). Thus, only the remaining two can be manipulated for the
compositions. To choose two out of the four, SVD provides a
straightforward method to compare the steady state behavior of various
first level control strategies.

In this example, DQ is the best first-level control scheme. It has a much

better condition number than the others. The second singular value of
this DQ scheme is much stronger than for the other schemes.

24
Application to Feedback Properties

25
26
Use of the minimum singular value of the plant: The minimum singular
value of the plant evaluated as a function of frequency is a useful measure for
evaluating the feasibility of achieving acceptable control. In general, we want
σ as large as possible.

Singular values for performance: In general, it is reasonable to require that

the gain e(ω ) 2 / r (ω ) 2 remains small for any direction of r (ω ) , including

the worst-case direction which gives a gin of σ ( S ( jω )) . Let 1/ wp ( jω )

represent the maximum allowable magnitude of e(ω ) 2 / r (ω ) 2 at each

frequency, This results in the following performance requirement:

, ∀ω ⇔ σ {wp S ( jω )} < 1, ∀ω ⇔ wp S ( jω
1
σ ( S ( jω )) < <1
wp ( jω ) ∞

Typical weight function is given asL

s / M + ωb
wp ( s) =
s + ωb A

which means 1/ wp equals A ≤ 1 at low frequencies, and equals to M ≥ 1

at high frequencies, and the asymptote crosses 1 at ωb (the bandwidth

frequency).

27