Chapter 21 ©

Goodwin, Graebe, Salgado , Prentice Hall 2000

Chapter 21

Exploiting SISO Techniques in

MIMO Control
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

In the case of SISO control, we found that one could use a

wide variety of synthesis methods. Some of these carry
over directly to the MIMO case. However, there are
several complexities that arise in MIMO situations. For
this reason, it is often desirable to use synthesis
procedures that are in some sense automated. This will be
the subject of the next few chapters. However, before we
delve into the full complexity of MIMO design, it is
appropriate that we pause to see when, if ever, SISO
techniques can be applied to MIMO problems directly.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We will study
 decentralized control as a mechanism for directly
exploiting SISO methods in a MIMO setting
 robustness issues associated with decentralized control .
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Completely Decentralized Control

Before we consider a fully interacting multivariable
design, it is often useful to check on whether a
completely decentralized design can achieve the
desired performance objectives. When applicable,
the advantage of a completely decentralized
controller, compared to a full MIMO controller, is
that it is simpler to understand, is easier to maintain,
and can be enhanced in a straightforward fashion (in
the case of a plant upgrade).
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Readers having previous exposure to practical control

will realize that a substantial proportion of real-world
systems will utilize decentralized architectures. Thus,
one is led to ask the question, is there ever a situation
in which decentralized control will not yield a
satisfactory solution? We will present several real-
world examples later in Chapter 22 that require MIMO
thinking to get a satisfactory solution. As a textbook
example of where decentralized control can break
down, consider the following MIMO example.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Example 21.1
Consider a two-input, two-output plant having the
transfer function
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Let us say that k12 and k21 depend on the operating

point (a common situation, in practice).
Operating point 1 (k12 = k21 = 0)
Clearly, there is no interaction at this operating point.
Thus, we can safely design two SISO controllers. To
be specific, say we aim for the following
complementary sensitivities:
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The corresponding controller transfer functions are

C1(s) and C2(s), where

The two independent loops perform as predicted by

the choice of complementary sensitivities.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Operating point 2 (k12 = k21 = 0.1)

We leave the controller as previously designed for
operating point 1. We apply a unit step in the
reference for output 1 at t = 1 and a unit step in the
reference for output 2 at t = 10. The closed-loop
response is shown on the next slide. These results
would probably be considered very acceptable, even
though the effects of coupling are now evident in the
response.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.1: Effects of weak interaction in control loop

with SISO design
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Operating point 3 (k12 = -1, k21 = 0.5)

With the same controllers and for the same test as
used at operating point 2, we obtain the results on the
next slide.
We see that a change in the reference in one loop
now affects the output in the other loop significantly.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.2: Effects of strong interaction in control

loops with SISO design
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Operating point 4 (k12 = -2, k21 = -1)

Now a simulation with the same reference signals
indicates that the whole system becomes unstable.
We see that the original SISO design has become
unacceptable at this final operating point.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Pairing of Inputs and Outputs

If one is to use a decentralized architecture, then one
needs to pair the inputs and outputs. In the case of
an m  m plant transfer function, there are m!
possible pairings. However, physical insight can
often be used to suggest sensible pairings.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Relative Gain Array

One method that can be used to suggest pairings is a
quantity known as the Relative Gain Array (RGA).
For a system with matrix transfer function Go(s), the
RGA is defined as a matrix  with the ijth element

where [Go(0)]ij and [Go-1(0)]ij denote the ijth element of

the plant d.c. gain and the jith element of the inverse of
the d.c. gain matrix respectively.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Note that [Go(0)]ij corresponds to the d.c. gain from

the ith input, ui, to the jth output, yj, while the rest of the
inputs, ul for l {1, 2, …, i-1, i+1, …, m} are kept
constant. Also [Go-1]ij is the reciprocal of the d.c. gain
from the ith input, ui, to the jth output, yj, while the rest
of the outputs, yl for l {1, 2, …, j-1, j+1, …, m} are
kept constant. Thus, the parameter ij provides an
indication of how sensible it is to pair the ith input with
the jth output.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

One usually aims to pick pairings such that the

diagonal entries of  are large. One also tries to
avoid pairings that result in negative diagonal entries
in .
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Example
Consider again the system

The RGA is then

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

For 1 > k12 > 0, 1 > k21 > 0, the RGA suggests the
pairing (u1, y1), (u2, y2). We recall from our earlier
study of this example that this pairing worked very
well for k12 = k21 = 0.1 and quite acceptably for k12
= -1, k21 = 0.5. In the latter case, the RGA is
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

However, for k12 = -2, k21 = -1 we found that the

centralized controller based on the pairing (u1, y1),
(u2, y2) was actually unstable. The corresponding
RGA in this case is

which indicates that we probably should have

changed to the pairing (u1, y2), (u2, y1).
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Example 21.3
Quadruple-tank apparatus.
Consider the quadruple-tank apparatus shown on the
next two slides.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We recall from Chapter 20 that this system has an

approximate transfer function,

The RGA for this system is

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

For 1 < 1 + 2 < 2, we recall from Chapter 20 that the

system is of minimum phase. If we take, for
example, 1 = 0.7 and 2 = 0.6, then the RGA is

This suggests that we can pair (u1, y1) and (u2, y2).
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Because the system is of minimum phase, the design

of a decentralized controller is relatively easy in this
case. For example, the following decentralized
controller gives the results shown on the next slide
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.3: Decentralized control of a minimum-

phase four-tank system
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

For 0 < 1 + 2 < 1, we recall from Chapter 20 that the

system is nonminimum phase. If we take, for example
1 = 0.43 and 2 = 0.34, then the system has a NMP
zero at s = 0.0229, and the relative gain array becomes
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

This suggests that (y1, y2) should be commuted for the

purposes of decentralized control,. This is physically
reasonable, given the flow patterns produced in this
case. This leads to a new RGA of
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Note, however, that control will still be much harder

than in the minimum-phase case. For example, the
following decentralized controllers give the results
shown on the next slide.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.4: Decentralized control of a nonminimum-

phase four-tank system
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Robustness Issues in Decentralized

Control
One way to carry out a decentralized control design
is to use a diagonal nominal model. The off-diagonal
terms then represent under-modelling, in the
terminology of Chapter 3.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Thus, say we have a model Go(s), then the nominal model for
decentralized control could be chosen as

and the additive model error would be

With this as a background, we can employ the robustness

checks described in Chapter 20. We recall that a sufficient
condition for robust stability is

where is the maximum singular value of

 ( G  1 ( j ) To ( j ))
G  1 ( j ) To ( j ).
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Example
Consider again the system
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

In this case, the various matrices arising in the

centralized design are
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The singular values, in this case, are simply the

magnitudes of the two off-diagonal elements. These
are plotted on the next slide for normalized values
k12 = k21 = 1.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.5: Singular Values of G1(j)To(j)

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We see that a sufficient condition for robust stability

of the decentralized control, with the pairing (u1, y1),
(u2, y2), is that |k12| < 1 and |k21| < 1. Observe that this
is conservative, but consistent with the performance
results presented earlier.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Example
Consider a MIMO system with

We first observe that the RGA for the nominal model

Go(s) is given by
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

This value of the RGA might lead to the hypothesis

that a correct pairing of inputs and outputs has been
made and that the interaction is weak. We thus
proceed to do a decentralized design leading to a
diagonal controller C(s) to achieve a complementary
sensitivity To(s), where
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

However, this controller, when applied to control the

full plant G(s), leads to closed-loop poles located at
-6.00, -2.49 ± j4.69, 0.23 ± j1.36, and -0.50 - an
unstable closed loop !
The lack of robustness in this example can be traced
to the fact that the required closed-loop bandwidth
includes a frequency range where the off-diagonal
frequency response is significant.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Feedforward Action in Decentralized

Control
Although it usually will not aid robust stability, the
performance of decentralized controllers is often
significantly enhanced by the judicious choice of
feedforward action to reduce coupling. Consider, for
example, the output response at port #1, i.e.

and, for simplicity, we consider only the effect of the jth

loop on the ith loop. We can then apply the feedforward
ideas developed in Chapter 10 to obtain the architecture
shown on the next slide.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.6: Feedforward action in decentralized control

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The feedforward gain G ffji (should

s) be chosen in such a way
that the coupling from the jth loop to the ith loop is compensated
in a particular, problem-dependent frequency band [0 ff] - i.e.

This can also be written as

from which we observe the necessity to build an inverse. Hence

all of the issues associated with building inverses discussed in
earlier chapters arise again.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Example 21.6
Consider again the system

with k12 = -1 and k21 = 0.5. We recall the results

presented earlier for this case.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

We see that there is little coupling from the first to the

second loop, but relatively strong coupling from the
second to the first loop. This suggests that
feedforward from the second input to the first loop
may be beneficial. To illustrate, we choose G ffji ( s ) to
completely compensate the coupling at d.c., i.e. G ffji ( s )
is chosen to be a constant G ff ( s )   , satisfying
ji
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The resulting modified MIMO system can be seen to

be modeled by

where
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The RGA is now  = diag(1, 1) and when we

redesign the decentralized controller, we obtain the
results presented on the next slide.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.7: Performance of a MIMO decentralized

control loop with interaction feedforward
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The above examples indicate that a little coupling

introduced into the controller can be quite helpful.
This, however, raises the question of how we can
systematically design coupled controllers that
rigorously take into account multivariable interaction.
This motivates us to study the latter topic, which will
be taken up in the next chapter. Before ending this
chapter, we investigate whether there exist simple
ways of converting an inherently MIMO problem to a
set of SISO problems.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Converting MIMO problems to

SISO Problems
Many MIMO problems can be modified so that
decentralized control becomes a more viable (or
attractive) option. For example, one can sometimes
use a precompensator to turn the resultant system into
a more nearly diagonal transfer function.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

To illustrate, say the nominal plant transfer function

is Go(s). If we introduce a precompensator P(s), then
the control loop appears as in the figure below.

Figure 21.8: Feedback control with plant

precompensation
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The design of Cp(s) can then be based on the

equivalent plant.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Several comments are in order regarding this strategy:

(i) A first attempt at designing P(s) might be to approximate Go(s)-
1
in some way. For example, one might use the d.c. gain matrix
Go(0)-1 as a precompensator, assuming this exists.
(ii) If dynamic precompensators are used, then one needs to check
that no unstable pole-zero cancellations are introduced between
the compensator and the original plant.
(iii) Various measures of resultant interactions can be introduced.
For example, the following terminology is frequently employed
in this context.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Dynamically decoupled
Dynamically decoupled: Here, every output depends
on one and only one input. The transfer-function matrix
H(s) is diagonal for all s. In this case, the problem
reduces to separate SISO control loops.
Band-decoupled and statically decoupled systems:
When the transfer-function matrix H(j) is diagonal
only in a finite frequency band, we say that the system
is decoupled in that band. In particular, we will say,
when H(0) is diagonal, that the system is statically
decoupled.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Triangularly coupled systems: A system is

triangularly coupled when the inputs and outputs can
be ordered in such a way that the transfer-function
matrix H(s) is either upper or lower triangular, for all
s. The coupling is then hierarchical.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Industrial Case Study

(Strip Flatness Control)
An an illustration of the use of simple
precompensators to convert a MIMO problem into
one in which SISO techniques can be employed, we
consider the problem of strip flatness control in
rolling mills. Actually, very similar issues arise in
many other problems including paper making and
plastic extrusion.
The next slide shows a typical rolling stand
configuration.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Typical rolling stand configuration

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

What is flatness in a Rolling Mill?

If rolling results in a nonuniform reduction of the
strip thickness across the strip width, then a residual
stress will be created, and buckling of the final
product may occur. A practical difficulty is that
flatness defects can be pulled out by the applied strip
tensions, so that they are not visible to the mill
operator. However, the buckling will become
apparent as the coil is unwound or after it is slit or cut
to length in subsequent processing operations.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Source of Flatness Problems

There are several sources of flatness problems, including the
following:

roll thermal cambers

incoming fed disturbances (profile, hardness, thickness)

transverse temperature gradients

roll stack deflections

incorrect ground roll cambers

roll wear

inappropriate mill setup (reduction, tension, force, roll bending)

lubrication effects.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

On the other hand, there are strong economic motives to

control strip flatness, including the following:

improved yield of prime-quality strip

increased throughput, due to faster permissible acceleration,
reduced threading delay, and higher rolling speed on shape-
critical products

more efficient recovery and operation on such downstream
units as annealing and continuous-process lines

reduced reprocessing of material on tension-leveling lines or
temper-rolling mills.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Control Options
In this context, there are several control options to
achieve improved flatness. These include roll tilt,
roll bending, and cooling sprays. These typically can
be separated by preprocessing the measured shape.
Here, we will focus on a particular aspect of the
cooling spray option. Note that flatness defects can
be measured across the strip by using a special
instrument called a Shape Meter. A typical control
configuration is shown on the next slide.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.9: Typical flatness-control set-up for rolling

mill
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

In this configuration, numerous cooling sprays are

located across the roll, and the flow through each
spray is controlled by a valve. The cool water
sprayed onto the roll reduces the thermal expansion.
The interesting thing is that each spray affects a large
section of the roll, not just the section directly
beneath it. This leads to an interactive MIMO
system, rather than a series of decoupled SISO
systems.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The thermal properties of the roll can be modeled

using basic laws of physics. This leads to a partial
differential equation, however, this can be discretized
to give a finite dimenional model. Such a model can
then be used as a calibration model to test control
system design strategies.
The main components of the heat flow inside a
typical roll are shown on the next slide.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Internal roll heat flows

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

For the purpose of control system design, it suffices

to use a simpler model. Such a model can be
developed by approximating the observed behavior
of the more complex calibration model. A key
feature of the observed behavior is that a single
cooling spray (one of the actuators) effects the radial
diameter of the roll and hence the measured strip
shape over a extended spatial area. This is
diagrammatically shown on the next slide.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Effect of a single spray on roll

diameter

Action of single spray

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Based on the above discussion, a simplified model

for this system (ignoring nonlinear heat-transfer
effects, etc.) is shown in the block diagram on the
next slide, where U denotes a vector of spray valve
positions and Y denotes the roll-thickness vector.
(The lines indicate vectors rather than single
signals).
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.10: Simplified flatness-control feedback loop

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The sprays affect the roll in a roughly exponential

fashion as described by the matrix M:

The parameter  represents the level of interactivity

in the system and is determined by the number of
sprays present and how close together they are.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

An interesting thing about this simplified model is that

the interaction is captured totally by the d.c. gain
matrix M. This suggests that we could design an
approximate precompensator by simply inverting this
matrix. This leads to
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Using this matrix to decouple the system has a nice

physical interpretation. Namely, it amounts to turning
off surrounding sprays when a spray is turned on. This
makes sense physically since we are preventing the
spread of the cooling effect by use of adjacent sprays.
The essential idea of the decoupling control strategy is
shown on the next slide.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Pictorial representation of coordinated

use of sprays to decouple control action

ith spray

(i-1)th spray (i+1)th spray

Nett shape
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

In summary, we can (approximately) decouple the

system simply by multiplying the control vector by the
appropriate inverse. This set-up is shown in the block
diagram on the next slide.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Figure 21.11: Flatness control with precompensation

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

The nominal decoupled system then becomes simply

H ( s )  diag ( s11 ) .With this new model, the controller can be
designed by using SISO methods. For example a set of
simple PI controllers linking each shape meter with the
corresponding spray would seem to suffice. (We assume that
the shape meters measure the shape of the rolls perfectly).
This idea is routinely used in this particular application and
leads to excellent results. (Of course, the practical problem
has many other features that we leave aside so as not to
distract from our key point here).
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Actually, control problems almost identical to the

above can be found in many alternative industrial
situations where there are longitudinal and traverse
effects. Examples are paper making and plastic
extrusion.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Strip flatness systems of the type (briefly) described

here are available commercially. The following
slides have been made from pamphlets describing a
commercial system sold by Industrial Automation
Services Pty. Ltd.
www.indauto.com.au
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Impact of MIMO Controller

Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

Summary
 A fundamental decision in MIMO synthesis pertains to the
choice of decentralized versus full MIMO control.
 Completely decentralized control

In completely decentralized control, the MIMO system is
approximated as a set of independent SISO systems

To do so, multivariable interactions are thought of as disturbances;
this is an approximation, because the interactions involve feedback,
whereas disturbance analysis actually presumes disturbances to be
independent inputs.

When applicable, the advantage of completely decentralized control
is that one can apply the simpler SISO theory.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000


Applicability of this approximation depends on the neglected
interaction dynamics, which can be viewed as modelilng errors;
robustness analysis can be applied to determine their impact.

Chances of success are increased by judiciously pairing inputs and
outputs (for example, by using the Relative Gain Array, RGA) and
by using feedforward.

Feedforward is often a very effective tool in MIMO problems.

Some MIMO problems can be better treated as SISO problems if a
precompensator is first used.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000

 There are several ways to quantify interactions in

multivariable systems, including their structure and their
strength.

Interactions can have a completely general structure (every input
potentially affects every output) or display particular patterns, such
as triangular or dominant diagonal; they can also display
frequency-dependent patterns, such as being statistically decoupled
or band-decoupled.

The lower the strength of interaction, the more nearly a system
behaves like a set of independent systems that can be analyzed and
controlled separately.
Chapter 21 ©
Goodwin, Graebe, Salgado , Prentice Hall 2000


Weak coupling can be due to the nature of the interacting dynamics
or to a separation in frequency range or time scale.

The stronger is the interaction, the more important it becomes to
view the multi-input multi-output system and its interactions as a
whole.

Compared to the SISO techniques discussed so far, viewing the
MIMO systems and its interactions as a whole requires generalized
synthesis and design techniques and insight. These will be the
topics of the following two chapters.