ECE5580: Multivariable Control in the Frequency Domain 6–1

Introduction to Multivariable Control

! In this chapter, we introduce the idea of multi-input multi-output

(MIMO) systems
! Specifically, we’ll develop the following ideas: transfer function
matrices, multivariable frequency response analysis, singular value
decomposition (SVD), the relative gain array (RGA), multivariable
control and multivariable zeros

Introduction

! Consider a MIMO plant having m inputs and l outputs

! Basic transfer function model is:
y.s/ D G.s/u.s/

where we have used bold-face symbols to make the vector-matrix

structure more apparant
! In general form, we can write
2 3 2 32 3
y1.s/ g11.s/ g12.s/ " " " g1m.s/ u1.s/
6 7 6 76 7
6 y2.s/ 7 6 g21.s/ g22.s/ " " " g2m.s/ 7 6 u2.s/ 7
6 7 6 76 7
6 ::: 7 D 6 ::: ::: ::: ::: 7 6 ::: 7
4 5 4 54 5
yl .s/ gl1.s/ gl2.s/ " " " glm.s/ um.s/
where it is easy to see that each of the elements of G.s/ is itself a
SISO transfer function gij .s/ relating input uj .s/ to output yi .s/

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–2

! A change in the first input u1 will generally affect all outputs,

y1; y2; : : : ; yl – this means there is interaction between inputs and
outputs
! The main difference between SISO and MIMO systems is the
existence of directions in MIMO systems – these will be explored in
greater detail in this chapter

Transfer function matrices for MIMO systems

! Transfer functions for MIMO systems are matrices as described

above, and as such must follow the rules of linear algebra
! Consider the following block diagrams:

+ v
u G1 y
+
G
u G1 G2 z
z
G2

Cascade system Positive feedback system

! C ASCADE RULE
For the cascade (series) interconnection of G1 and G2 , the overall
transfer function matrix is G D G2G1.
– Note that the order of the transfer function matrices is reversed
from how they appear

! F EEDBACK RULE
With reference to the positive feedback system, we have
v D .I $ L/$1u where L D G2G1 is the transfer function around the
loop.
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–3
! P USH - THROUGH RULE
For matrices of appropriate dimension,
G1 .I $ G2G1/$1 D .I $ G1G2/$1 G1

! From these rules, we may state the following MIMO RULE:

Start from the output and write down the blocks as you meet them
when moving backwards (against signal flow) towards the input. If
you exit from a feedback loop then inlude a term .I $ L/$1 for positive
feedback (or .I C L/$1for negative feedback) where L is the loop
transfer function. Parallel branches should be treated independently
and their contributions added together.

Example 6.1

! Find the transfer function from w to z in the block diagram below.

P22

+
P21 K P12
+

+
w P11 z
+

S OLUTION
! Start at the output z and move backward towards w.

– One branch beyond the summing junction goes to P11 directly

c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–4
– At the other branch, continue backward to encounter P12 and then
K
! Exit the feedback loop and get term .I $ L/$1(positive feedback) with
L D P22K
! Finally, encounter term P21
! The complete transfer function is given by:
! "
z D P11 C P12K .I $ P22K/$1 P21 w

Negative feedback control systems

! Consider the negative feedback system shown below.
d2 d1

+ u+ + + +
r K G y
−

! We define L to be the loop transfer function when breaking the loop

at the plant output,

L D GK

! Accordingly, the sensitivity and complementary sensitivity are defined

as
S , .I C L/$1
T , I $ S D L .I C L/$1
– To make it clear we’re defining these transfer function at the plant
output, we further define
LO D L SO D S TO D T
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–5

! We next define LI to be the loop transfer function at the input to the

plant,
LI D KG

! We then define
SI , .I C LI /$1
TI , I $ SI D LI .I C LI /$1

! Some useful relationships follow:

.I C L/$1 C .I C L/$1 L D S C T D I
G .I C KG/$1 D .I C GK/$1 G
GK .I C GK/$1 D G .I C KG/$1 K D .I C GK/$1 GK
! "$1
L .I C L/$1 D I C L$1 D .I C L/$1 L D T

– A simple rule to remember: “G comes first and then G and K

alternate in sequence”.

Multivariable frequency response analysis

! G.s/ is a transfer function matrix which is a function of the complex

variable s

– For s D s0, G.s0/ becomes a constant complex-valued matrix

! G.j!/ is a complex matrix function representing the response to

sinusoidal signals of frequency !

Obtaining the frequency response from G .s/

! Consider the system G.s/ with input vector d.s/ and output vector
y.s/:
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–6

d G(s) y

– Input-output dynamics are given by

y.s/ D G.s/d.s/
! A sinusoidal input to channel j is given by
! "
dj .t/ D dj 0 sin !t C ˛j

! The output in channel i is a sinusoid with the same frequency

yi .t/ D yi0 sin .!t C ˇi /
– Amplification gain is given by:
yi0 ˇ ˇ
D gij .j!/ˇ
ˇ
dj 0
– Phase shift is given by:
ˇi $ ˛j D †gij .j!/
! We can represent the result in phasor notation as:
yi .!/ D gij .j!/ dj .!/

where dj .!/ D dj 0e j˛j and yi .!/ D yi0e jˇi .

– In matrix form, we have
y .!/ D G .j!/ d .!/

Example 6.2
! Consider a 2 % 2 multivariable system where we simultaneously apply
sinusoidal signals of the same frequency ! to the two input channels:
" # " #
d1.t/ d10 sin .!t C ˛1/
d.t/ D D
d2.t/ d20 sin .!t C ˛2/

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–7
! The output signal,
" # " #
y1.t / y10 sin .!t C ˇ1/
y.t/ D D
y2.t / y20 sin .!t C ˇ2/
can be computed by multiplying the complex matrix G .j!/ by the
complex vector d .!/:

y .!/ D G .j!/ d .!/

where " # " #

jˇ1 j˛1
y10e d10e
y .!/ D and d .!/ D
y20e jˇ2 d20e j˛2

Directions in multivariable systems

! For a SISO system, y D Gd , the gain is given by

jy .!/j jG .j!/ d .!/j
D D jG .j!/j
jd .!/j jd .!/j
– The gain depends on !, but is otherwise independent of the
magnitude jd .!/j

! For MIMO systems, things are very different, because input and
output are vectors

– So we need to use a measure appropriate for vectors – norm!

! Using the vector 2-norm, we have

sX
ˇ ˇ2 q 2
kd .!/k2 D ˇdj .!/ˇ D d C d 2 C " " "
10 20
j

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–8
sX q
ky .!/k2 D jyi .!/j2 D 2
y10 2
C y20 C """
i

! The gain of the system transfer function matrix G.s/ is

ky .!/k2 kG.j!/d.!/k2
D
kd .!/k2 kd.!/k2
! In the MIMO case, the gain depends on frequency !, but is
independent of the norm kd .!/k2
– However, the gain depends on the direction of the input vector d
! The MIMO gain given above is recognized as the induced 2-norm –
its maximum is computed as the maximum singular value of G:
kGdk2
max D max kGdk2 D !N .G/
d¤0 kdk2 kdk2 D1

– Conversely, the minimum gain is the minimum singular value of G:

kGdk2
min D min kGdk2 D ! .G/
d¤0 kdk2 kdk2 D1

Example 6.3
! Consider the five different inputs shown below (kdk2 D 1):
" # " # " #
1 0 0:7071
d1 D ; d2 D ; d3 D
0 1 0:7071
" # " #
0:7071 0:6
d4 D ; d5 D
$0:7071 $0:8
! For the 2 % 2 system
" #
5 4
G1 D
3 2

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–9

! The five inputs d j lead to the following corresponding output vectors:

" # " # " #
5 4 6:36
y1 D ; y2 D ; y3 D
3 2 3:54
" # " #
0:7071 $0:2
y4 D ; y5 D
0:7071 0:2
! The corresponding 2-norms are:

ky 1k2 D 5:83

ky 2k2 D 4:47

ky 3k2 D 7:30

ky 4k2 D 1:00

ky 5k2 D 0:28

! The figure below shows a functional plot of the matrix gain as a

function of a parameterized input direction

– It is easy to see that maximum and minimum gains are achieved at

distinct input vector directions

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–10

σ = 7.34

σ = 0.27

! Another way to visualize the singular value decomposition is

presented below where the domain and corresponding image of the
input and output map are shown.

Eigenvalues are a poor measure of gain

! Consider the system y D Gd with

" #
0 100
GD
0 0

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–11
– Note that both eigenvalues are equal to zero, but the gain is equal
to 100

! P ROBLEM : Eigenvalues measure the gain for the special case when
the inputs and the outputs are in the same direction (that of the
eigenvectors)
! For generalizations of jGj when G is a matrix, we need the concept of
a matrix norm, denoted kGk
! Two important properties of norms:

– Triangle inequality
kG1 C G2k & kG1k C kG2k
– Multiplicative property
kG1G2k & kG1k " kG2k

! Note that the spectral radius

" .G/ , j#max .G/j

does not satisfy the properties of a matrix norm

Singular value decomposition

! Any matrix G may be decomposed into its singular value

decomposition,
G D U˙V '

– ˙ is an l % m matrix with k D min fl; mg non-negative singular

values !i arranged in descending order along its main diagonal,
where p
!i .G/ D #i .G 'G/

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–12

– U is an l % l unitary matrix of output singular vectors, ui

– V is an m % m unitary matrix of input singular vectors, vi

Example 6.4

! The SVD of a real 2 % 2 matrix can always be written in the form

" #" #" #T
cos $1 $ sin $1 !1 0 cos $2 ˙ sin $2
GD
sin $1 cos $1 0 !2 $ sin $2 ˙ cos $2
– Note that U and V involve rotations and their columns are
orthonormal

! The singular values are sometimes called the principal values or

principal gains, and the associated directions are called principal
directions
Input and output directions

! The column vectors of U , denoted by ui , represent the output

directions of the plant

– They are orthogonal and of unit length (orthonormal), that is

kui k2 D 1

u'i ui D 1

ui uj D 0; i ¤j

! The column vectors of V , denoted by vi , are orthogonal and of unit

length, and represent the input directions.
! Input and output directions are related through the singular values,
Gvi D !i ui
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–13

– If we consider an input in the direction vi ; then the output is in the

direction ui

! Since kvi k2 D 1 and kui k2 D 1, !i gives the gain of the matrix G in this
direction
kGvi k2
!i .G/ D kGvi k2 D
kvi k2
Maximum and minimum singular values

! The largest gain for any input direction is equal to the maximum
singular value
kGdk2 kGv1k2
!N .G/ D !1 .G/ D max D
d¤0 kdk2 kv1k2
! The smallest gain for any input direction is
kGdk2 kGvk k2
! .G/ D !k .G/ D min D
d¤0 kdk2 kvk k2
n o
where k D min l; m
! Therefore, for any vector d we have the general relationship
kGdk2
! .G/ & & !N .G/
kdk2
! Define u1 D u,
N and v1 D v,
N

– Then it follows,
G vN D !N uN
– vN is the maximum principal input direction, and corresponds to the
input direction with the largest amplification
– uN is the maximum principal output direction, and is the output
direction in which the inputs are most effective
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–14

! Next define uk D u , and vk D v

– Then we can write
Gv D !u
– v is the minimum principal input direction, and corresponds to the
input direction with the smallest amplification
– u is the minimum principal output direction, and is the output
direction in which the inputs are least effective
! These important vectors are associated with the “highest gain” and
’lowest gain” directions respectively

Example 6.5
! Consider the system of Example 6.3,
" #
5 4
G1 D
3 2
! The singular value decomposition of G1 is
" #" #" #'
0:872 0:490 7:343 0 0:794 $0:608
G1 D
0:490 $0:872 0 0:272 0:608 0:794
– The largest gain of 7:343 is for an input in the direction
" #
0:794
vN D
0:608
– The smallest gain of 0:272 is for an input in the direction
" #
$0:608
vD
0:794
! Note from the form of G1 that both inputs affect both outputs – thus
the system exhibits interaction
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–15
! Note also that the system is ill-conditioned, meaning that some
combinations of inputs have a strong effect on the outputs, whereas
others have a weak effect
!N .G1/ 7:343
– C ONDITION N UMBER: % .G1/ D D D 27:0
! .G1/ 0:272
Example 6.6: Distillation Process

! Consider the following steady-state model of a distillation column:

" #
87:8 $86:4
GD
108:2 $109:6
– Since the element magnitudes are much larger than 1, there
should be no problem with input constraints – however, the
low-gain direction could be problematic
" #" #" #'
0:625 $0:781 197:2 0 0:707 $0:708
GD
$:781 0:625 0 1:39 $0:708 $0:707

– The distillation process is clearly ill-conditioned, with condition

number % .G/ D 197:2=1:39 D 141:7
– The system is such that the two inputs counteract each other in the
direction of v in steady-state

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–16
Singular values for performance

! The maximum singular value is very useful in terms of frequency

domain performance and robustness
! For SISO systems we found that jS .j!/j D je.!/j=jr.!/j evaluated over
frequency gives useful information for feedback control
! Generalizations for MIMO systems can be obtained if we consider the
ratio ke.!/k2=kr.!/k2 where we write
ke .!/k2
! .S .j!// & & !N .S .j!//
kr .!/k2
! For performance we want the gain ke.!/k2=kr.!/k2 small for any direction
of r .!/

1
!N .S .j!// < ; 8!
jwP .j!/j
, !N .wP S / < 1; 8!
, kwP S k1 < 1

where the H1 norm is the peak of the maximum singular value of the
frequency response:
kM .s/k1 , max !N .M .j!//
!

! The singular values of S .j!/ may be plotted versus frequency –

typically they are small at low frequencies where feedback is effective,
and approach 1 at high frequencies because any real system is
strictly proper
p
! B ANDWIDTH, !B : frequency where !N .S/ crosses 1= 2 D 0:7071 from
below
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–17

– The bandwidth is at least !B for any direction of the input signal

– Since S D .I C L/$1, the singular value inequality
1
! .A/ $ 1 & & ! .A/ C 1
!N .I C A/$1
gives
1
! .L/ $ 1 & & ! .L/ C 1
!N .S/
– So at low frequencies,
1
! .L/ ( 1 ) !N .S/ )
! .L/
– And at high frequencies,
!N .L/ * 1 ) !N .S/ ) 1

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–18
Relative Gain Array (RGA)

! The RGA is a useful technique applied to multivariable systems to

assess interaction
! Mathematically, the RGA of a non-singular square complex matrix G
is defined as
! $1"T
RGA .G/ D & .G/ , G % G

where here ’%’ denotes the Hadamard product (element-by-element

multiplication)

– In Matlab we write RGA = G.*pinv(G).’

! The RGA of a transfer matrix is normally computed as a function of

frequency

RGA as an interaction measure

! Let uj and y i denote a particular input-output pair for plant G.s/ – we

wish to use uj to control y i
! There will be two extreme cases:

– All other loops are open: uk D 0; 8k ¤ j

– All other loops are closed with perfect control: yk D 0; 8k ¤ i

! Evaluating gain @y i=@uj for the two cases

# $
@y i
– Other loops open: D gij
@uj uk D0; k¤j
# $
@y i
– Other loops closed: , gO ij
@uj y D0; k¤i
k

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–19

! Here, gij D ŒG'ij is the ij th element of G, whereas gO ij is the inverse of

the j i th element of G $1:
1
gO ij D
ŒG $1'j i
! The ratio between the “loops open” and “loops closed” gains turns out
to be a useful measure of interaction – we define the ij th relative gain
as
gij % &
#ij , D ŒG'ij G $1 j i
gO ij
– The RGA is the corresponding matrix of these relative gains

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–20

Example 6.7: RGA for 2 % 2 system

! Consider the system given by
" # " #" #
y1 g11 g12 u1
D
y2 g21 g22 u2
where our task is to control ouput y1 using input u1.
! For the ouputs, we write
y1 D g11u1 C g12u2
y2 D g21u1 C g22u2

! We consider two cases:

– Case 1: u2 D 0
y1 D g11u1
– Case 2: y2 D 0
g21
u2 D $ u1
g22
! Substituting, we get
# $
g21
y1 D g11 $ g12 u1
g22

y2 D 0

where we define
g21
gO 11 D g11 $ g22
g22

! The interpretation here is that the gain changes from g11 to gO 11 as we

close the other loop
! Practically speaking, we prefer to pair input-output varables so that
the corresponding #ij is close to 1
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–21
Example 6.8

! Consider a blending process where we mix sugar (u1 ) and water (u2)
to make a given amount of a soft drink (y1 D F ) with a given sugar
fraction (y2 D x)
! Mass balance gives:
F1 C F2 D F
F1 D xF

! Linearization yields (note that the process has no dynamics):

dF1 C dF2 D dF
dF1 D x 'dF C F 'dx

! Assigning u1 D dF1, u2 D dF2, y1 D dF , and y2 D dx gives the model,

y1 D u1 C u2
# $
1 $ x' x'
y2 D u1 $ ' u2
F' F
where x ' D 0:2 is the nominal steady-state sugar fraction and
F ' D 2 kg=s is the nominal amount
! We the obtain the corresponding transfer function matrix,
2 3 " #
1 1 1 1
G.s/ D 4 1 $ x ' x ' 5D
$ 0:4 $0:1
F' F'
! Constructing the RGA we obtain (at all frequencies)
" # " #
' '
x 1$x 0:2 0:8
&D D
1 $ x' x' 0:8 0:2

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–22

! For de-coupled control, we should pair on the off-diagonal elements;

i.e., use u1 to control y1.

– Physically, this corresponds to using the largest stream (water, u2)

to control the amount (y1 D F )

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–23

Example 6.9: Frequency-dependent RGA

! The following model describes a large pressurized vessel (e.g., as
used in offshore oil-gas separations)
– Inputs are: (i) valve positions for liquid (u1), and (ii) vapor (u2) flow
– Outputs are: (i) liquid volume (y1), and (ii) pressure (y2)
! Transfer function matrix, G.s/ D
" #
$5s
0:01e $34:54 .s C :0572/ 1:913
! $4
"
.s C 1:72 10$4/ .4:32s C 1/ $30:22s $9:188 s C 6:95 10
! The RGA matrix depends on frequency
– At steady-state (s D 0), we see that g21 D 0; hence ƒ.0/ D I ,
suggesting diagonal pairing should be used
– However, examining other frequencies we see that off-diagonal
terms can approach 1

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–24

Introduction to MIMO robustness

! An example is presented to motivate the need for a deeper

understanding of robustness

Example 6.10: Spinning satellite

! Consider the following model of angular velocity control of a satellite

spinning about one of its principal axes:
" #
2
1 s$a a .s C 1/
G.s/ D 2 I a D 10
s C a2 $a .s C 1/ s $ a2
– A minimal state-space realization is given by:
2 3
" # 6 0 a 1 07
A B 6 $a 0 0 1 7
D6 6
7
7
C D 4 1 a 0 05
$a 1 0 0
! The plant has a pair of j!-axis poles at s D ˙ja , so it needs to be
stabilized
! For stabilization, try the simple diagonal constant controller K D I ;
this gives
" #
1 1 a
T .s/ D GK .I C GK/$1 D
s C 1 $a 1
! N OMINAL STABILITY. Two closed-loop poles are located at s D $1, so
it is stable

– Examining the closed-loop system,

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–25

" # " # " #

0 a 1 a $1 0
Acl D A $ BKC D $ D
$a 0 $a 1 0 $1
! N OMINAL PERFORMANCE. The frequency-dependent singular values
of L D GK D G are shown below.

– Note that ! .L/ D 1 at low frequencies and starts dropping off

around ! D 10 rad=sec
– Since ! .L/ never exceeds 1 , we do not have tight control in the
low-gain direction
– Also, the large off-diagonal elements in T .s/ show that we have
strong interaction in the closed-loop system

! R OBUST S TABILITY. Check stability one loop at a time – here we have

broken the loop at the first input

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–26
z1 w1

+
+ − G
−

! Loop transfer function –

z1 1
, L1.s/ D ) GM D 1; PM D 90ı
w1 s
! Good robustness? Let’s check.

– Consider input gain uncertainty, and let (1 and (2 denote the

relative gain error in each input channel.
0 0
u1 D .1 C (1/ u1; u2 D .1 C (2/ u2
– The resulting change in the state-space B-matrix is
" #
1 C (1 0
B0 D
0 1 C (2
– The corresponding closed-loop A-matrix is
" # " #" #
0 a 1 C (1 0 1 a
A0cl D A $ BKC D $
$a 0 0 1 C (2 $a 1
– And characteristic polynomial –

! "
det dI $ A0cl D s 2 C .2 C (1 C (2/ s

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–27
! 2 "
C1 C (1 C (2 C a C 1 (1(2
– Stability conditions are:
$1 < (1 < 1
(2 D 0
and
$1 < (2 < 1
(1 D 0
– But only small simultaneous changes in the two channels: e.g., let
(1 D $(2, then the system is unstable (a0 < 0) for
1
j(1j > p ) 0:1
a C1
2

! Summary:

– Checking single-loop margins is inadequate for MIMO problems

– Large values of !.T
N / or !.S/
N indicate robustness problems

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–28
Characteristic Loci

Introduction
! Recall that for a square system,
z .s/
det ŒG .s/' D ˛
p .s/
where z.s/ and p.s/ are the zero and pole polynomial of G.s/ after the
determinant has been adjusted to have p.s/ as its denominator.
! For a unity feedback system, the return difference matrix is defined as
F .s/ D I C G .s/ K .s/ D I C L .s/

! In terms of the closed-loop transfer function matrix relating r to y , we

have
T .s/ D .I C L .s//$1 L .s/ D L .s/ .I C L .s//$1 D L .s/ F .s/$1

! This expression for T .s/ gives a multivariable generalization of the

scalar relationship,
L .s/
t .s/ D D L .s/ .1 C L .s//$1 D L .s/ .1 C L .s//$1
1 C L .s/
! We’ve seen that the closed-loop poles are determined by the poles of
.I C L.s//$1and hence the zeros of I C L.s/ D F .s/
– Consider the free response of the closed-loop output, y.s/, (with
r.s/ D 0)
y.s/ D L.s/e.s/ D $L.s/y.s/

.I C L.s// y.s/ D F .s/y.s/ D 0

– Clearly this expression is only satisfied for values of s for which
det fF .s/g D 0; thus, closed-loop poles are described by the zeros
of F .s/
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–29

– Since the poles of F .s/ are the same as the poles of L.s/; we may
write
pc .s/ closed loop characteristic polynomial
det fF .s/g D D
po.s/ open loop characteristic polynomial

! Consider the scalar relationships,

n.s/
g.s/ D
d.s/

g.s/ n.s/
t.s/ D D
1 C g.s/ n.s/ C d.s/
! Then the open- and closed-loop polynomials, po.s/ D d.s/, and
pc .s/ D n.s/ C d.s/ satisfy
pc .s/ n.s/ C d.s/ n.s/
D D1C D 1 C g.s/ D f .s/
po.s/ d.s/ d.s/
– Intuitively we expect det fF .s/g to be the vehicle for the
generalization of the Nyquist criterion to the MIMO case

! We define the C HARACTERISTIC G AIN F UNCTION, g.s/ of G.s/ to be

the natural extension of the graphical Nyquist criterion

– The eigenfunctions of G.s/; denoted gi .s/, define the branches of

g.s/ and form the generalization of the scalar loop gain g.s/
– As such, g.s/ obeys the fame fundamental open- to closed-loop
relationship, namely ti .s/ D gi .s/=1Cgi .s/

! The frequency response plots of the gi .s/; gi .j!/, are called the
“characteristic loci” of G.s/

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–30

! Using these functions, multivariable analysis and design can be

carried out employing well-established SISO techniques of
gain-phase margins, bandwidth, dc gain, etc....

Physical motivation

! Extending the idea of scalar transfer functions to transfer matrices, we

have for a 2 % 2 system,
y1 D g11 .j!/ u1 C g12 .j!/ u2

y2 D g21 .j!/ u1 C g22 .j!/ u2

– In general, each component of u is stretched and rotated by a

different amount – so it is not possible to use any arbitrary input
vector phasor to define a multivariable gain
– We need a special input phasor u D w that produces an output
phasor y that is a scalar multiple of the input phasor, i.e.,
G .j!/ w D gw
– Thus the concept of gain we are after emerges as the n
eigenvalues of G.j!/; gi .j!/ for i D 1; 2; : : : ; n
– The corresponding phasors are the eigenvectors wi .j!/
associated with the gi .j!/

! Repeated eigenvalue calculations at a number of frequency points

result in n plots called the characteristic loci (CL) of G.s/

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–31
Example 6.11

! Consider a two-tank system with transfer function matrix

2 3
3 1:5
6 sC3 sC3 7
6 7
6
G.s/ D 6 7
7
4 1 $1 5
sC6 sC6
! To calculate the CL, we select values of s D j!, compute G.j!/,
perform an eigenvalue-vector decomposition and plot in polar form:

Open- and closed-loop characteristic loci

! We may derive closed-loop characteristic loci directly from the

open-loop form by considering the spectral decomposition of G.j!/
and T .j!/

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–32

! It is easy to show that under unity feedback,

G .j!/ D W .j!/ ƒG .j!/ V .j!/

T .j!/ D W .j!/ ƒT .j!/ V .j!/

and that
ƒT .j!/ D .I C ƒG .j!//$1
2 3
g1 .j!/
6 1 C g1 .j!/ 7
6 7
6 g2 .j!/ 7
6 7
D66 1 C g2 .j!/ 7
7
6 ::: 7
6 7
4 gn .j!/ 5
1 C gn .j!/
! This gives rise to two important results:

– The closed-loop CL, ti .j!/, are related to the open-loop CL,

gi .j!/, in the same simple way as for SISO systems
– The eigenframes of G.j!/ remain invariant under unity feedback

! A corollary to this is that a MIMO system is stable under unity

feedback if its open-loop CL satisfy the classical Nyquist stability
criterion

Characteristic gain and the characteristic locus

Example 6.12

! Consider the following transfer function matrix,

" #
10 3s C 4 $ .4s C 6/
G .s/ D
.s C 1/ .s C 2/ 2s C 3 $ .3s C 5/

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–33

! Analytically computing the eigenfunctions, we get

det fgI $ G.s/g D 0

10 100
D g2 C g$ D0
.s C 1/ .s C 2/ .s C 1/ .s C 2/
! Solving gives the two eigenfunctions,
10
g1.s/ D
sC2
$10
g2.s/ D
sC1
! Computing the eigenvector functions, we obtain

" # " #
g1.s/ $ g22.s/ 20s C 30 2
w1.s/ D D
g21.s/ .s C 1/ .s C 2/ 1
" # " #
g2.s/ $ g22.s/ 20s C 30 1
w2.s/ D D
g21.s/ .s C 1/ .s C 2/ 1
which gives,
" #
2 1
W D
1 1

! The system of Example 6.12 was especially constructed to give

eigenfunctions which turn out to be rational functions of s – but this is
not the case in general
! Indeed, if we perturb the matrix just slightly, we get a different result:
" #
10 3s C 4 C ( $ .4s C 6/
G 0 .s/ D
.s C 1/ .s C 2/ 2s C 3 $ .3s C 5/
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–34

– Solving for the eigenfunctions,

det fgI $ G.s/g D 0

10 C 10( 100 $ ( .3s C 5/
D g2 C g$ D0
.s C 1/ .s C 2/ 2
.s C 1/ .s C 2/ 2

– So that,
$ .10 C 10(/
g1;2.s/ D
2 .s C 1/2 .s C 2/2
q% &
.10 C 10(/2 C 4 .100 $ ( .3s C 5// .s C 1/2 .s C 2/2
˙
2 .s C 1/2 .s C 2/2
! The discriminant (under the square root) is no longer a perfect square
– thus the equation is irreducible over the field of rational functions

– Because of this, g defined on the complex plane is multi-valued –

i.e., closed curves on the s-plane may not map to closed curves on
the complex g-plane
– A full treatment of this effect requires the introdution of Riemann
surfaces and will not be addressed here

Generalized Nyquist Stability Criterion

! The forgoing development leads to a multivariable generalization of

the SISO Nyquist criterion and may be stated as follows:

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–35

A multivariable system is closed-loop stable (under unity

feedback) if and only if the net sum of critical point encir-
clements by the CL of the transfer function matrix G.s/ is
in a counterclockwise sense and is equal to the number of
open-loop unstable poles. For open-loop stable systems
the net sum of critical point encirclements must be zero.

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–36
Multivariable interaction

! In general a signal being applied to the i th input will cause all the
outputs (j D i and j ¤ i ) to respond – this defines interaction
! It is generally desirable to suppress interaction so that the i th input
excites mostly the i th output while the response of the j ¤ i inputs
remain below acceptable levels
! In principal, it is possible to supress interaction by ensuring the
moduli of the CL be sufficienly large, because then ti ) 1

– But this cannot normally be done due to stability concerns

! An alternative approach involves the characteristic direction

(eigenvector function) set

– Low interaction requires the standard basis directions e i to be

reproduced at the outputs – which demands that vector e i should
be an eigenvector of T .s/ , or equivalently of L.s/
– Thus a sufficient condition for low interaction is that the
characteristic direction set of L.j!/ approach the standard basis
set
– One may take as a convenient measure of interaction the angle
between the vectors wk .j!/ and the standard basis directions e i
ˇ ' ˇ
ˇw .j!/ e i ˇ
cos $i .j!/ D k
wk '.j!/w'k .j!/
where the $i are termed the “misalignment angles”.
ı The convention is adopted that k is chosen so that wk is the
eigenvector that makes the smallest angle $i with e i at
frequency !
c 2016, M. Scott Trimboli
Lecture notes prepared by M. Scott Trimboli. Copyright #
ECE5580, Introduction to Multivariable Control 6–37
! In summary,

– Interaction is determined at low frequencies by $i and/or moduli of

the CL of the open-loop operator
– Interaction is determined at high frequencies by the $i

c 2016, M. Scott Trimboli

Lecture notes prepared by M. Scott Trimboli. Copyright #
(mostly blank)
(mostly blank)
(mostly blank)