1.

Introduction

Prior to the 1940s, most industrial systems were run essentially manually
or using on-off control. Many operators were needed to keep watch on the
many variables in the plant. With the continuous industrialization over the
last several decades, manufacturing and production have taken off at an in-
credibly high speed in almost every part of the world. As a consequence of
the expanding scale and volume of production, there could be hundreds or
even thousands of variables to be controlled in a plant. The manual effort
thus needed in operation is tremendous. With increasing labor and equip-
ment costs, and with eager demand for high precision, quality and efficiency,
the idea of employing operators for the control of physical systems rapidly
became uneconomical and infeasible. Automatic control thus becomes a solu-
tion much sought after. The fundamental component in an automatic control
system is the so called “controller”. It could either be a piece of hardware
or software code in a computer. Its job is to receive information about the
system from a variety of sensors, process it, and automatically generate com-
mands for corrective action to bring the variable of interest to its desired
value or trajectory. Wide applications of automatic control have driven great
attention to its theoretical development. Since the age of the familiar clas-
sical control theory, many new and sophisticated theories have evolved and
striking developments have taken place especially since 1950s. Optimal and
stochastic control, state-space theory and adaptive control are only a few to
name. The advances in microelectronics which give cheap microprocessors
whose computational power is continuously increasing have also acted as a
supportive catalyst in the development of new concepts by providing conve-
nient test grounds and platforms for extensive simulations to be performed
in the various stages of design and verification, and enabling effective im-
plementation of advanced control algorithms. Today, automatic controllers
have found a much wider scope of usage than only in industrial processes. It
may range from missile tracking in military applications, down to water and
temperature control in washing machines. Automatic control systems have
permeated life in all advanced societies.

Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 1-9, 2003.

Springer-Verlag Berlin Heidelberg 2003
2 1. Introduction

1.1 Multivariable Systems and Decoupling Control

A typical system (natural or man-made) will have several variables to be con-

trolled and is called multivariable systems if so. Multivariable systems can
be found almost everywhere. In the office, the temperature and humidity are
crucial to comfort. Water flow speed and rate are two key measures of a river.
A robot needs six degree-of-freedoms in order to have a full range of posi-
tioning, and the same can be said to airplanes and missiles. Some phenomena
are unique to multivariable systems only, and could not occur in single vari-
able systems. For instance, a multivariable system can have a pole and zero
which coincide with each other and yet do not cancel each other. A zero
of some element in a multivariable system play no role in the system prop-
erties. The most important feature with a multivariable system is possible
cross-couplings or interactions between its variables, i.e., one input variable
may affect all the output variables. They prevent the control engineer from
designing each loop independently as adjusting controller parameters of one
loop affects the performance of another, sometimes to the extent of desta-
bilizing the entire system. It is the multivariable interactions that accounts
for essential difference for design methodologies between single-variable and
multivariable control systems. In general, multivariable control is much more
difficult than single variable control, and it is the topic of interest in the
present book.
A multivariable system will be simplified to a number of single variable
systems if it has no cross-couplings between variables, and is called decoupled
in this case. As mentioned above, a given multivariable plant usually has cou-
plings. A design strategy is then to design a multivariable controller which
can decouple the plant, that is, the resulting control system has no more cou-
plings between the desired reference variables and the plant output variables.
It should be stressed that decoulping control is popular NOT mainly because
it can simplify multivariable control system design, BUT rather because it is a
desired feature in many practical applications. Firstly, decoupling is required
for ease of system operations, at least in process and chemical industry, as
otherwise, technicians operating a multivariable control system can hardly
decide the values of multiple set-points to meet their target. Secondly, poor
decoupling could be the principal common control problem in industry, see
the recent survey (Kong, 1995) from the leading controller manufactures such
as Fish-Rosemount, Yokogawa and Foxboro. According to our own industrial
experience, poor decoupling in closed-loop could, in many cases, lead to poor
diagonal loop performance. Conversely, good decoupling is helpful for good
loop performance. An evidence is that some optimal control without any
decoupling requirement could eventually result in a decoupled closed-loop
(Dickman and Sivan, 1985; Linnemann and Wang, 1993). Thirdly, it should
be noted that even if the decoupling requirement is relaxed to the limited cou-
pling, it will still lead to almost decoupling if the number of inputs/outputs
is large. Roughly, if the number is 10, total couplings from all other loops
 1.1 Multivariable Systems and Decoupling Control 3

are limited to 30%, then each off-diagonal element will have relative gain less
than 3%, so the system is almost decoupled.
It should be, however, pointed out that there are cases where decoupling
should not be used, instead, couplings are deliberately employed to boost
performance such as airplane control systems. The extreme case is that de-
coupling cannot be used at all since it may destabilize the system. In short,
decoupling is a common practice in industry, but needs to be justified from
the real situation and requirements before use.
The study of decoupling linear time-invariant multivariable systems has
received considerable attentions in both control theory and industrial prac-
tice for several decades, dating back at least to (Voznesenskii, 1938). Dur-
ing that early times, this problem was treated with transfer matrices, see
(Kavanagh, 1966; Tsien, 1954) for reviews. The state space approach to de-
coupling was initiated by Morgan (Morgan, 1964). A seminal paper by Falb
and Wolovich (Falb and Wolovich, 1967) presented a necessary and suffi-
cient condition on the solvability of decoupling problem for square systems
described by state space models. The equivalent condition for systems de-
scribed by transfer matrices was given by Gilbert (Gilbert, 1969) later. In
these investigations, the problem is confined to the case of scalar input and
output blocks and hence with equal number of inputs and outputs. The more
general block decoupling problem was first defined and solved by Wonham
and Morse (Morse and Wonham, 1970; Wonham and Morse, 1970) using ge-
ometric approach, see also the monograph (Wonham, 1986). For this kind
of general block decoupling problem, an alternative approach was developed
by Silverman and Payne (Silverman and Payne, 1971) based on structure
algorithm proposed by Silverman (Silverman, 1969). In these works, it is
still required that the transformation matrix between the new control in-
puts and old control inputs is nonsingular. The complete solution to gen-
eral Morgan problem, without any assumptions on either system matrices
or feedback matrices, is due to Descusse et al. (Descusse, 1991; Descusse et
al., 1988). Other developments can be found in (Commault et al., 1991; Des-
oer and Gundes, 1986; Hautus and Heymann, 1983; Koussiouris, 1979; Kous-
siouris, 1980; Pernebo, 1981; Williams and Antsaklis, 1986), to name a few.
A parallel avenue for the study of decoupling problem is for unity out-
put feedback systems based on input-output models. The existence condition
for decoupling by unity output feedback compensation is that the plant has
full row rank, which is very simple and well known. But the problem of de-
coupling in such a configuration while maintaining internal stability of the
decoupled system appears to be much more difficult. Under the assump-
tion that the plant has no unstable poles coinciding with its zeros, the di-
agonal decoupling problem is solvable and a parameterization of controllers
solving the problem could be obtained (Safonov and Chen, 1982; Vardu-
lakis, 1987; Peng, 1990; Desoer, 1990; Lin and Hsieh, 1991). The assumption
is however not necessary and is relaxed later by Linnemann and Maier (1990)
for 2 × 2 plants, and by Wang (1992) for square plants, along with a con-
4 1. Introduction

troller parameterization. Alternative necessary and sufficient conditions for

this problem was also derived based on transfer matrices and residue by Lin
(1997) and on the new notion of the minimal C + -decoupler by Wang and
Yang (2002). The results of Wang (1992) are generalized to the block de-
coupling problem for possibly non-square plants and the performance lim-
itations in the decoupled system are compared with those of non-decoupled
systems in Linnemann and Wang (1993).
The works cited above consider delay-free systems only. Time delay is
however present and popular in process and chemical industries, and causes
a serious obstacle to good process operation and control. It prevents high gain
of a conventional controller from being used, leading to offset and sluggish
system response. The problem is usually further complicated by multivariable
nature of most plants in operation. Input-output loops in a multivariable
plant usually have different time delays, and for a particular loop its output
could be affected by all the inputs through likely different time delays. As
a result, such a plant may then be represented by a multivariable transfer
function matrix having multiple (or different) time delays around a operating
point. Wang et al. (1997a) presented a transfer function matrix approach to
decoupling design. It is shown there that the controller diagonal elements
can be first designed with respect to the “equivalent processes” only and the
controller off-diagonal elements then determined to achieve decoupling. A
systematic method was proposed in Zhang (1999) for the decoupling design of
general multivariable controller for processes with multi-delays. A controller
of lowest complexity is found so as to achieve fastest loops with acceptable
overshoot and minimum loop interactions.
For single variable processes with delay, Smith (1957) suggested a compen-
sation scheme which can remove the delay from the closed-loop characteristic
equation and thus eases feedback control design and improves set-point re-
sponse greatly. Like the SISO case, multivariable delay compensation may
be devised in hope of performance enhancement. This was done for multi-
variable systems with single delay (Alevisakis and Seborg, 1973; Alevisakis
and Seborg, 1974) and for multi-delays (Ogunnaike and Ray, 1979; Ogun-
naike et al., 1983). In the perfect model-process match case, they can, like
SISO Smith’s, eliminate all the time delays from the system characteristic
equation but, unlike SISO Smith’s, fails to achieve the desired output perfor-
mance expected from the delay-free design because the delays in the process
could mix up the outputs of delay-free part to generate messy actual output
response. Thus, the performance of this MIMO Smith might sometimes be
poor (Garcia and Morari, 1985). Jerome and Ray (1986) proposed a different
version of multivariable Smith predictor control to improve overall perfor-
mance. However, their design of the primary controller is based on a transfer
matrix with time delays and is difficult to carry out. This also contradicts the
Smith’s philosophy of delay removal from closed-loop characteristic equation
for ease of stable design. The problem was solved in Wang et al. (2000) by
incorporating decoupling into the MIMO Smith scheme.
 1.2 Disturbance Decoupling 5

It is well known that a Smith predictor controller can be put into an

equivalent internal model control (IMC) structure. The IMC was introduced
by Garcia and Morari (1982) and studied thoroughly in Morari and Zafiriou
(1989) as a powerful control design strategy for linear systems. In principle, if
a multivariable delay model, Ĝ(s), is factorized as Ĝ(s) = Ĝ+ (s)Ĝ− (s) such
that Ĝ+ (s) is diagonal and contains all the time delays and non-minimum
phase zeros of Ĝ(s), then the IMC controller can be designed, like the scalar
case, as K(s) = {Ĝ− (s)}−1 with a possible filter appended to it, and the per-
formance improvement of the resultant IMC controller over various versions
of the multivariable Smith predictor schemes can be expected (Garcia and
Morari, 1985). For a scalar transfer function, it is quite easy to obtain Ĝ+ (s).
However, it becomes much more difficult for a transfer matrix with multi-
delays. The factorization is affected not only by the time delays in individual
elements but also by their distributions within the transfer function matrix,
and a non-minimum phase zero is not related to that of elements of the trans-
fer matrix at all (Holt and Morari, 1985b; Holt and Morari, 1985a). In Zhang
(1999), an effective approach to the IMC analysis and design is proposed for
decoupling and stabilizing linear stable square multivariable processes with
multi-delays. There, the characteristics of all the controllers which solve this
decoupling problem with stability and the resulting closed-loop systems will
be given in terms of their unavoidable time delays and non-minimum phase
zeros which actually determine performance limitations for IMC systems.
It can be concluded that decoupling control is now well developed, at least
for nominal design, and it is timely and desirable to write a book which is
devoted to such an important topic and gives a comprehensive and in-depth
coverage of the topic. In the author’s opinion, the emphasis of future research
should be on robust decoupling and in particular, practical and stable design
techniques which can achieve the specified coupling constraints under the
given plant uncertainty.

1.2 Disturbance Decoupling

It is well known (Astrom and Hagglund, 1995) that the attenuation of load
disturbance is of a primary concern for any control system design, and is even
the ultimate objective for process control, where the set-point may be kept
unchanged for years (Luyben, 1990). In fact, countermeasure of disturbance
is one of the top factors for successful and failed applications (Takatsu and
Itoh, 1999). If the disturbance is measurable, feedforward control is a useful
tool for cancelling its effect on the system output (Ogata, 1997), while feed-
back control may still be used as usual for stability and robustness. Formally,
the problem of (dynamic) disturbance decoupling is to find a compensator
such that the resulting closed-loop transfer function from the disturbance
to the controlled output is equal to zero. This problem is well-known and
6 1. Introduction

has been investigated extensively. It was actually the starting point for the
development of a geometric approach to systems theory. Numerous investiga-
tors, employing a variety of mathematical formulations and techniques, have
studied this problem.
In a geometric setting the problem using a state feedback was consid-
ered by Basile and Marro (1969) and Wonham (1986). Equivalent frequency
domain solvability conditions were given by Bhattacharyya (1980), Bhat-
tacharyya (1982). It is more realistic to assume that only output is available
for feedback. In this context, the problem was solved by Akashi and Imai
(1979) and Schumacher (1980). However, they did not consider the issue
of stability. The disturbance decoupling by output feedback with stability
or pole placement was solved for the first time by Willems and Commault
(1981). This problem was also considered from an input-output viewpoint by
Kucera (1983) for single variable system and by Pernebo (1981) for multi-
variable systems. The structure of the control system which decouples the
disturbances is now well understood in both geometric terms and frequency
domain terms.
To reject unmeasurable disturbances which are more often encountered in
industry, one possible way is to rely on the single controller in the feedback
system in addition to normal performance requirements. Then, there will be
inevitably a design trade-off between the set-point response and disturbance
rejection performance (Astrom and Hagglund, 1995). To alleviate this prob-
lem, a control scheme which introduces some add-on mechanism to the con-
ventional feedback system was introduced, first in the filed of the servo control
in mechatronics, and is called the disturbance observer (Ohnishi, 1987). It
was further refined by Umeno and Hori (1991). The disturbance observer
estimates the equivalent disturbance as the difference between the actual
process output and the output of the nominal model. The estimate is then
fed to a process inverse model to cancel the disturbance effect on the output.
The disturbance observer makes use of the plant inverse, and its applica-
tion needs some approximation as the plant inverse is usually not physically
un-realizable due to improperness and/or time delay.
The complete disturbance decoupling is usually difficult to achieve. A
more realistic requirement will be static or asymptotic disturbance decou-
pling. The latter problem falls in a general framework of asymptotic tracking
and regulation problem, a fundamental control system problem. Wonham
(1986) gives a complete exposition, where solutions together with (necessary
and sufficient) solvability conditions, and a procedure for the construction
of a compensator are obtained. This problem had been alternatively been
considered by Davison (1975) in an algebraic setting. Both works are carried
out in a state-space viewpoint. This problem, which can also be formulated
in a frequency domain setting, was also approached from an input-output
viewpoint such as Francis (1977) who considered the case where feedback
signals coincide with the outputs; Chen and Pearson (1978) who considered
the case where feedback signals are certain functions of outputs; and finally
 1.2 Disturbance Decoupling 7

Pernebo (1981) who considered the general case where feedback signals may
be different from outputs and not related to each other. The crucial condition
for the solvability in frequency domain setting is the existence condition for
the solution of Diophantine equation (Kucera, 1979), which, in a special case,
is simplified to skew-primeness (Wolovich, 1978). Now, the problem has been
well solved in both state space and frequency domains. The relevant solutions
in terms of the internal model principle are popular in control society.
In practice, we often encounter the situation where the reference com-
mands to be tracked and/or disturbance to be rejected are periodic signals,
e.g., repetitive commands or operations for mechanical systems such as in-
dustrial robots, and disturbances depending on the frequency of the power
supply (Hara et al., 1988). Disturbances acting on the track-following servo
system of an optical disk drive inherently contain significant periodic compo-
nents that cause tracking errors of a periodic nature. For such disturbances,
the controller designed for step type reference tracking and/or disturbance
rejection will inevitably give an uncompensated error. Though the internal
model principle seems applicable, the unstable modes of a periodic distur-
bance is infinite and no longer rational, and actually it contains a time delay
term in its denominator. For such delay and unstable systems, no stabilizing
scheme could be found in the literature. Therefore, one has to abandon the in-
ternal model principle and seek other solutions. Hara et al. (1988) pioneered a
method for periodic disturbance rejection, which is now called repetitive con-
trol. However, this method potentially makes the closed-loop system prone
to instability, because the internal positive feedback loop that generates a
periodic signal reduces the stability margin. Consequently, the trade-off be-
tween system stability and disturbance rejection is an important yet difficult
aspect of the repetitive control system design. Moon et al. (1998) proposed
a repetitive controller design method for periodic disturbance rejection with
uncertain plant coefficients. The design is performed by analyzing the fre-
quency domain properties, and Nyquist plots play a central role throughout
the design phase. It is noted that such a method is robust at the expense
of performance deterioration compared with the nominal case, and further,
the disturbance cannot be fully compensated for even under the ideal case.
Another way to solve the periodic disturbance problem is to use the double
controller scheme (Tian and Gao, 1998). However, the complexity and the
lack of tuning rules hinder its application. A plug-in adaptive controller (Hu
and Tomizuka, 1993; Miyamoto et al., 1999) is proposed to reject periodic
disturbances. An appealing feature is that turning on or off the plug module
will not affect the original structure. However, the shortcoming of this method
is that the analysis and implementations are somewhat more complex than
the conventional model based algorithm. An alternative scheme, called the
virtual feedforward control, is proposed in Zhang (2000) for asymptotic re-
jection of periodic disturbance. The periodic disturbance is estimated when
a steady state periodic error is detected, and the virtual feedforward control
is then activated to compensate for such a disturbance.
8 1. Introduction

1.3 Organization of the Book

The book assumes the pre-requisite of basic linear system theory from readers.
It is organized as follows.
Chapter 2 reviews some notions and results on linear systems. Multivari-
able linear dynamic systems and their representations are introduced. Poly-
nomial and rational function matrices are studied in details. Matrix fraction
descriptions and multivariable pole and zeros are covered. General methods
for model reduction are highlighted and two special algorithms taking care
of stability of reduced models are presented. Popular formulas for conver-
sions between continuous time and discrete time systems are discussed and
a technique with high accuracy and stability is described.
Chapter 3 considers stability and robustness of linear feedback systems.
Internal stability of general interconnected systems is addressed and a pow-
erful stability condition is derived which is applicable to systems with any
feedback and/or feedforward combinations. A special attention is paid to the
conventional unity output feedback configuration and the simplifying condi-
tions and Nyquist-like criteria are given. The plant uncertainties and feedback
system robustness are introduced.
Chapter 4 considers decoupling problem by state feedback. Necessary and
sufficient solvability conditions are derived and formulas for calculating feed-
back and feedforward gain matrices given. Pole and zeros of decoupled sys-
tems are discussed. Geometric methods are not covered.
Chapter 5 addresses decoupling problem by unity output feedback com-
pensation. The polynomial matrix approach is adopted. The diagonal de-
coupling problem for square plants is first solved and then extended to the
general block decoupling case for non-square plants. A unified and indepen-
dent solution is also presented. In all the cases, stability is included in the
discussion, the necessary and sufficient condition for solvability is given, and
the set of all the compensators solving the problem is characterized.
Chapter 6 discusses decoupling problem for plants with time delay. Con-
ventional unity output feedback configuration is used exclusively. Our empha-
sis is to develop decoupling methodologies with necessary theoretical supports
as well as the controller design details for possible practical applications. The
new decoupling equations are derived in a transfer function matrix setting,
and achievable performance after decoupling is analyzed where characteriza-
tion of the unavoidable time delays and non-minimum phase zeros that are
inherent in a feedback loop is given. An effective decoupling control design
method is then presented with stability and robustness analysis.
Chapter 7 considers decoupling problem in connection with time delay
compensation. The presence of time delay in a feedback control loop could
impose serious control performance limitations, and they can be released
only by time delay compensation. The Smith control scheme and the internal
model control (IMC) structure are popular for SISO time delay compensation,
and are extended to MIMO delay systems in this chapter.
 1.3 Organization of the Book 9

Chapter 8 addresses near-decoupling problem. In most applications, exact

decoupling is not necessary or impossible. The concept of near-decoupling
control is then proposed. Design methods for near-decoupling controllers are
presented for four general (block-decoupling) cases: (i) exact models and state
feedback, (ii) uncertain models and state feedback, (iii) exact models and
dynamic output feedback, and (iv) uncertain models and dynamic output
feedback. All theses results are given in the form of linear matrix inequalities,
which makes the numerical solutions easily tractable.
Chapter 9 deals with decoupling a disturbance from the plant output
in dynamic sense. For measurable disturbances, the feedforward compensa-
tion scheme is employed, and necessary and sufficient conditions as well as
controller parameterization are given. For unmeasurable disturbances, the
disturbance observer scheme is introduced, and its theory and design are
presented. Stability is a crucial issue and is covered in both schemes.
Chapter 10 is concerned with static or asymptotic disturbance decoupling.
For general disturbances, the internal model principle is developed and the
relevant design and computations are discussed. For a specific yet common
case of periodic disturbances, the virtual feedforward control is proposed. It
is an add-on function on top of normal feedback control. It is shown that the
closed-loop stability is not affected by the proposed control and thus there is
no design trade-off between disturbance rejection and stability.
The relations among the chapters are simple: Chapters 4-10 are almost
independent of each other, and they may require some knowledge of linear
systems and robust control covered in Chapters 2 and 3. For instance, Chapter
5 makes use of polynomial and rational matrix theory (Sections 2 and 3 of
Chapter 2) as well as stability conditions on unity output feedback systems
Section 2 of Chapter 3); Chapters 6 and 7 need model reduction (Section
4 of Chapter 2) and stability and robustness analysis (Sections 2 and 3 of
Chapter 3).
2. Representations of Linear Dynamic Systems

In this chapter, we will review some preliminary notions and results on linear
systems that will be needed subsequently for the solution of various decou-
pling problems in later chapters. Multivariable linear dynamic systems and
their representations are introduced. Polynomial and rational function ma-
trices are studied in details. Matrix fraction descriptions and multivariable
poles and zeros are covered. State-space realizations from matrix fraction
descriptions are discussed. Model reduction is important in many ways, ex-
isting methods are highlighted and two algorithms taking care of stability
of reduced models are presented. Conversions between continuous time and
discrete time systems are reviewed and a technique with high accuracy and
stability preservation is described.

2.1 Linear Dynamical Systems

Let a finite dimensional linear time-invariant dynamical system be described

by the following linear constant coefficient equations:
ẋ(t) = Ax(t) + Bu(t), x(t0 ) = x0 , (2.1)
y(t) = Cx(t) + Du(t), (2.2)
where x(t) ∈ Rn is called the system state, x(t0 ) the initial condition of the
system, u(t) ∈ Rm the system input, and y(t) ∈ Rp the system output. The
A, B, C and D are real constant matrices with appropriate dimensions. A
dynamical system with m = 1 and p = 1 is known as a single-input and
single-output (SISO) system, otherwise as an multiple-input and multiple-
output (MIMO) system. In this book, we shall be concerned mainly with
MIMO systems.
For the given initial condition x(t0 ) and the input u(t), the dynamical
system will have its solution, or response x(t) and y(t) for t ≥ t0 , which can
be obtained from the following formulas:
Z t
x(t) = eA(t−t0 ) x(t0 ) + eA(t−τ ) Bu(τ )dτ, (2.3)
t0
y(t) = Cx(t) + Du(t). (2.4)

Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 11-64, 2003.

Springer-Verlag Berlin Heidelberg 2003
12 2. Representations of Linear Dynamic Systems

The matrix exponential eAt is defined by

1 2 2 1
eAt = I + At + A t + A3 t3 + · · · (2.5)
2! 3!
and it can also be calculated by

eAt = L−1 {(sI − A)−1 }, (2.6)

where L and L−1 stand for the Laplace and inverse Laplace transforms,
respectively.
In the case of u(t) = 0, ∀t ≥ t0 , it is easy to see from the solution (2.3)
that for any t1 ≥ t0 and t ≥ t0 , we have

x(t) = eA(t−t1 ) x(t1 ).

Therefore, the matrix function Φ(t, t1 ) = eA(t−t1 ) acts as a transformation

from one state to another, and thus Φ(t, t1 ) is usually called the state tran-
sition matrix. Since the state of a linear system at one time can be obtained
from the state at another time through the transition matrix, we can assume
without loss of generality that t0 = 0. This will be assumed in the sequel.

Definition 2.1.1. (Controllability) The dynamical system described by

(2.1) or the pair (A, B) is said to be controllable if there exists an input
u(τ ), 0 ≤ τ ≤ t1 , which transfers the initial state x(0) = x0 to any desired
state x(t1 ) = x1 .

Theorem 2.1.1. The system (2.1) is controllable if and only if one of the
following conditions is true:
i) The controllability matrix:
 
Qc := B AB A2 B · · · An−1 B ,

has full row rank (over the real number field R);
ii) For α ∈ Rn , αT eAt B = 0 implies α = 0 for any t > 0, i.e, the n rows of
eAt B are linearly independent (over the real number field R);
iii) The matrix [sI − A, B] has full row rank for any s ∈ C, the complex
number field;
iv) The eigenvalues of A − BK can be freely assigned (with the restriction
that complex eigenvalues are in conjugate pairs) by a suitable choice of
K.

The observability is a dual concept to controllability, and it can be tested

by Theorem 2.1.1 with the substitution of (A, B) = (AT , C T ).

Definition 2.1.2. The dynamical system (2.1) is said to be (internally) sta-

ble if all the eigenvalues λ(A) of A are in the open left half plane, i.e,
Re λ(A) < 0. A matrix A with such a property is said to be stable or Hurwitz.
 2.1 Linear Dynamical Systems 13

Note that the unforced solution x(t) to the stable system (2.1), i.e., under
u(t) ≡ 0, meets x(t) = eAt x(0) → 0 when t → ∞.

Definition 2.1.3. The dynamical system (2.1), or the pair (A, B), is said
to be state-feedback stabilizable if there exists a state feedback u = −Kx such
that the feedback system,

ẋ = (A − Bk)x,

is stable.

Theorem 2.1.2. The system in (2.1) is state-feedback stabilize if and only

if the matrix [sI − A, B] has full row rank for all Re s ≥ 0.

Taking the Laplace transform of the system in (2.1) and (2.2) under the
zero initial condition x(0) = 0 yields the transfer function matrix of the
system:

G(s) = C(sI − A)−1 B + D. (2.7)

On the other hand, given the transfer function matrix G(s), the system in
(2.1) and (2.2) is said to be a (state-space) realization of G(s) if (2.7) holds.
The realization is called minimal if (A, B) is controllable and (A, C) is ob-
servable.
The computation of the transfer matrix G(s) from the state-space model
can be carried out with the Faddeev algorithm as follows. Let the character-
istic polynomial of the n × n matrix A be denoted by

det(sI − A) := φ(s) = sn + αn−1 sn−1 + αn−2 sn−2 + · · · + α0 . (2.8)

Then, write
1
(sI − A)−1 = (Γn−1 sn−1 + Γn−2 sn−2 + · · · + Γ0 ), (2.9)
φ(s)
where αn−1 , αn−2 , . . . , α0 , Γn−1 , Γn−2 , · · · , Γ0 can be calculated in a recursive
manner from
Γn−1 = I, αn−1 = −tr (AΓn−1 ),
Γn−2 = AΓn−1 + αn−1 I, αn−2 = −tr (AΓn−2 )/2,
Γn−3 = AΓn−2 + αn−2 I, αn−3 = −tr (AΓn−3 )/3,
··· ···
Γi = AΓi+1 + αi+1 I, αi = −tr (AΓi )/(n − i),
··· ···
Γ0 = AΓ1 + α1 I, α0 = −tr (AΓ0 )/n,
0 = AΓ0 + α0 I,
where tr (X), the trace of X, is the sum of all the diagonal elements of the
matrix X.
14 2. Representations of Linear Dynamic Systems

Example 2.1.1. Consider the following multivariable state-space system:

   
−1 −2 0 01
ẋ(t) =  −3 −4 0  x(t) +  1 0  u(t),
0 0 −5 00
   
101 10
y(t) = x(t) + u(t).
001 01
Using the algorithm shown above, we compute

Γ2 s2 + Γ1 s + Γ0
(sI − A)−1 = ,
φ(s)

where
     
100 9 −2 0 20 −10 0
Γ2 =  0 1 0  , Γ1 =  −3 6 0  , Γ0 =  −15 5 0  ,
001 0 0 5 0 0 −2

and
φ(s) = s3 + 10s2 + 23s − 10,
which in turn lead to
1
C(sI − A)−1 B = {CΓ2 Bs2 + CΓ1 Bs + CΓ0 B}
φ(s)
     
1 01 2 −2 9 −10 20
= s + s+
φ(s) 00 0 0 0 0
 2

1 −2s − 10 s + 9s + 20
= ,
φ(s) 0 0
and
G(s) = C(sI − A)−1 B + D
" #
−2s−10 s2 +9s+20
+ 1
= φ(s) φ(s) . ♦
0 1

The transfer function matrices for MIMO systems without time delay
are in fact rational function matrices of complex variable, namely, matrices
whose generic element is a rational function, a ratio of polynomials with real
coefficients. A transfer function matrix is said to be proper when each element
is a proper rational function, i.e.,

lim G(s) = K < ∞,

s→∞

where the notation K < ∞ means that each element of matrix K is finite.
Analogously, G(s) is strictly proper if
 2.2 Polynomial Matrices 15

lim G(s) = 0.
s→∞

A rational matrix G(s) is said to be stable or system (2.1) is said to be input-

output stable if G(s) is proper and all elements of G(s), whenever expressed
as ratio of polynomials without common roots, have their poles in the open
left half plane only.

2.2 Polynomial Matrices

As one has seen in the proceeding section, a multivariable state-space model
is a collection of coupled first-order differential equations. But more naturally,
the model one gets from first-principle modelling and/or system identification
is usually a set of coupled differential equations whose order may be greater
than one. A elective circuit described by

ÿ1 (t) + 3ẏ1 (t) + 2y2 (t) = 2ü1 (t) + u2 (t) + 3u̇2 (t),
ẏ2 (t) + 3y1 (t) = u1 (t) + 2u2 (t),

is such an example. Taking the Laplace transform under zero initial condition
yields

D(s)Y (s) = N (s)U (s),

where Y (s) = [Y1 (s) Y2 (s)]T , U (s) = [U1 (s) U2 (s)]T ;

 2   2 
s + 3s 2 2s 3s + 1
D(s) = and N (s) =
s 3 1 2
are polynomial matrices. The transfer function matrix of the system is

G(s) = D−1 (s)N (s),

which is a rational function matrix expressed as a fraction of two polynomial

matrices.
Therefore, a study of polynomial and rational function matrices is im-
portant as it is a basic building block in the analysis and synthesis of mul-
tivariable systems. We will discuss polynomial matrices in this section while
rational matrices are the topic of the next section. We begin with the intro-
duction of some standard notations.
Let R be the field of real numbers, R[s] the ring of polynomials with
coefficients in R and R(s) the field of fractions over R[s]:
n n(s) o
R(s) := t(s)|t(s) = , n(s), d(s) ∈ R[s], d(s) 6≡ 0 .
d(s)

R(s) is called the field of (real) rational functions. A matrix P (s) whose
elements are polynomials is called a polynomial matrix. The set of p × m
16 2. Representations of Linear Dynamic Systems

polynomial matrices is denoted by Rp×m [s]. Similarly, Let Rp×m (s) denote
the set of p × m matrices with elements in R(s). A matrix T (s) ∈ Rp×m (s)
is called a (real) rational function matrix. The rank of a polynomial or ratio-
nal function matrix is the maximum number of its linearly independent row
vectors or its column vectors over the field R(s) and it is denoted by rank
T (s).

2.2.1 Unimodular Matrices and Elementary Operations

Let P (s) ∈ Rm×m [s] be a square polynomial matrix. In general, its determi-
nant, denoted by det P (s), is a polynomial. The following two special cases
deserve attention.

Definition 2.2.1. P (s) is called singular if detP (s) = 0; otherwise, it is

called nonsingular; P (s) is called unimodular if detP (s) = α, where α is a
non-zero real number.

Note that if P(s) is nonsingular, then its inverse P −1 (s) exists and is in
general a rational function matrix. In the case of unimodular P (s), however,
the inversion formula, P −1 (s) = adj(P (s))/ det P (s), tells us that P −1 (s) be-
comes a polynomial matrix, and is also unimodular due to
det P (s) det P −1 (s) = 1. In fact, P (s) is unimodular if and only if P (s) and
P −1 (s) are both polynomial matrices.
Unimodular matrices are closely related to elementary operations. To be
precise, elementary row and column operations on any polynomial matrix
P (s) ∈ Rp×m [s] are defined as follows:
i) interchange any two rows or columns of P(s);
ii) multiply row or column i of P(s) by a non-zero real number;
iii) add to row or column i of P(s) a polynomial multiple of row or column
j, j 6= i.
These elementary operations can be accomplished by multiplying the given
P (s) on the left or on the right by elementary unimodular matrices, namely
matrices obtained by performing the above elementary operations on the
identity matrix I. It can also be shown that every unimodular matrix may
be represented as the product of a finite number of elementary matrices.

2.2.2 Degree and Division

For a polynomial p(s), we denote its degree by ∂(p(s)) or ∂p(s) in case of no

confusion. Recall that

(i) for a(s), b(s) ∈ R[s] such that a(s)b(s) 6= 0, there holds ∂(a(s)b(s)) ≥
∂(a(s));
 2.2 Polynomial Matrices 17

(ii) the polynomial division theorem: for every a(s), b(s) ∈ R[s], b(s) 6= 0,
there exists two elements q(s), r(s) ∈ R[s] such that a(s) = q(s)b(s)+r(s)
and either r(s) = 0 or ∂r(s) < ∂b(s).
We now extend these facts to the matrix case.
Consider a p × m polynomial matrix P(s). The i − th column (resp.
row) degree of P(s) is defined as the highest degree of all m polynomials
in the i − th column (resp. row) of P (s), and denoted by ∂ci (P (s)) (resp.
∂ri (P (s))). Let ∂ci P (s) = µi , i = 1, 2, . . . , m; ∂ri P (s) = νi , i = 1, 2, . . . , p;
Sc (s) = diag{sµ1 , sµ2 , · · · , sµm } and Sr (s) = diag{sν1 , sν2 , · · · , sνp }. P (s)
can be expressed by
P (s) = Phc Sc (s) + Lc (s)
= Sr (s)Phr + Lr (s),
where Phc (resp. Phr ) is the highest column (resp. row) degree coefficient
matrix of P (s), and Lc (s) (resp. Lr (s)) has strictly lower column (resp. row)
degrees than ∂ci P (s) (resp. ∂ri P (s)).
In the case of square matrices, p = m, it readily follows from linear algebra
that
det P (s) = det Phc · sµ + lower − degree terms in s
and
det P (s) = sν det Phr + lower − degree terms in s
m
X m
X
with µ = µi and ν = νi , which imply, respectively
i=1 i=1

m
X
∂ detP (s) ≤ µi ,
i=1

and
m
X
∂ detP (s) ≤ νi .
i=1

The equality holds if Phc (resp. Phr ) is nonsingular.

Definition 2.2.2. A polynomial matrix is called column (resp. row)-reduced

if Phc (resp. Phr ) is nonsingular.

Example 2.2.1. Let

 2 
s + 1 s2
P (s) = .
2s s

One writes
18 2. Representations of Linear Dynamic Systems

    
11 s2 0 1 0
+
P (s) = 00 0 s2 2s s
Phc Sc (s) Lc (s)
    
s2 0 11 10
+ .
= 0 s 21 00
Sr (s) Phr Lr (s)

P (s) is row-reduced since det Phr 6= 0 but not column-reduced since

det Phc = 0. ♦
By post- (resp. pre-) multiplication by an appropriate unimodular matrix,
it is possible to reduce nonsingular P (s) to a column (resp. row)-reduced form.
Proposition 2.2.1. For a nonsingular polynomial matrix P (s), there are
unimodular matrices UR (s) and UL (s) such that

P (s) = P (s)UR (s)

is column-reduced, and

Pe(s) = UL (s)P (s)

is row-reduced.
Continue Example 2.2.1, where P (s) is not column-reduced. To reduce
column degrees, we examine just the highest-degree terms in each column:
 2 2
s s
,
0 0
which can be transformed to
 2
0s
0 0
by a post-multiplication with
 
1 0
UR (s) = .
−1 1
Thus, we perform
    2
s2 + 1 s2 1 0 1s
P (s)UR (s) = =
2s s −1 1 s s
    
01 s 0 10
+ ,
= 10 0 s2 0s
P̄hc S̄c (s) L̄c (s)
which is already column-reduced.
 2.2 Polynomial Matrices 19

Consider now two polynomial matrices N (s) ∈ Rp×m [s] and D(s) ∈
m×m
R [s] with D(s) nonsingular. In general, N (s)D−1 (s) will be a ratio-
nal function matrix. By carrying out polynomial divisions in each element of
N (s)D−1 (s), we have

N (s)D−1 (s) = Q(s) + W (s),

where Q(s) ∈ Rp×m [s] is polynomial and W (s) ∈ Rp×m (s) is strictly proper.
It can be re-written as

N (s) = Q(s)D(s) + R(s), (2.10)

where R(s) := W (s)D(s) = N (s) − Q(s)D(s) is polynomial and R(s)D−1 (s) =

W (s) is strictly proper as mentioned before. Such Q(s) and R(s) are unique
since if Q̄(s) and R̄(s) is another pair such that (2.10) holds and R̄(s)D(s)−1
is strictly proper, then from N (s) = Q̄(s)D(s) + R̄(s) and (2.10), it follows
that
h i
Q(s) − Q̄(s) = R̄(s) − R(s) D−1 (s), (2.11)

where the left-hand side is polynomial and the right-hand side is strictly
proper. For (2.11) to hold true, both sides must be zero. Hence Q(s) = Q̄(s)
and R(s) = R̄(s).

Proposition 2.2.2. Let D(s) ∈ Rm×m [s] be nonsingular. Then, for any
N (s) ∈ Rp×m [s] there exist unique Q(s), R(s) ∈ Rp×m [s] such that (2.10)
holds true with R(s)D(s)−1 ∈ Rp×m (s) strictly proper.

Analogously, given a nonsingular A(s) ∈ Rp×p [s], then for any B(s) ∈
p×m
R [s] there exist unique Q(s), R(s) ∈ Rp×m [s] such that

B(s) = A(s)Q(s) + R(s), (2.12)

with A(s)−1 R(s) ∈ Rp×m (s) being strictly proper.

Example 2.2.2. Let

 
  s 1
s 1
D(s) = , N (s) =  0 s2  .
−1 1
−1 0

Given that
 
−1 1 1 −1
D (s) = ,
s+1 1 s

we obtain
20 2. Representations of Linear Dynamic Systems

     
s+1 s 1 0 0 0
1  s2 s3  =  s − 1 s2 − s + 1  + 1  1 −1  .
N (s)D−1 (s) =
s+1 s+1
−1 1 0 0 −1 1

Hence,
     
1 0 0 0   0 0
1 s 1
Q(s) =  s − 1 s2 − s + 1  , R(s) =  1 −1  =  1 0 . ♦
s+1 −1 1
0 0 −1 1 −1 0

We now present an algorithm for direct determination of Q(s) and R(s).

This method is the most natural generalization of the scalar polynomial di-
vision algorithm. To begin, let D(s) be column-reduced with column de-
grees ∂ci D(s) = µi , i = 1, 2, · · · , m. Let ∂ci N (s) = νi , i = 1, 2, · · · , m,
µ = max{µi , i = 1, 2, · · · , m}, l = µ + max{νi − µi , i = 1, 2, · · · , m}, and
Sc (s) = diag{sµ1 , sµ2 , · · · , sµm }. Then D(s), N (s) and R(s) can be written
as
D(s) = [Dµ + Dµ−1 s−1 + · · · + D1 s−µ+1 + D0 s−µ ]Sc (s), (2.13)
l−µ l−µ−1 −µ+1 −µ
N (s) = [Nl s + Nl−1 s + · · · + N1 s + N0 s ]Sc (s), (2.14)
R(s) = [Rµ−1 s−1 + Rµ−2 s−2 + · · · + R1 s−µ+1 + R0 s−µ ]Sc (s), (2.15)
where Dµ is nonsingular, and the columns of Di , Ni and Ri corresponding to
the negative powers of s are zero. Finally, let Q(s) be

Q(s) = Qγ sγ + Qγ−1 sγ−1 + · · · + Q1 s + Q0 , (2.16)

where γ, Qi as well as Ri are to be determined.

Substituting (2.13)-(2.16) into (2.10) and equating the terms of the same
degree on its both sides yield
µ
X
Ni = Qi−j Dj + Ri , i = γ + µ, γ + µ − 1, · · · , 2, 1, 0,
j=0

" µ−1
#
X
Qi−µ = Ni − Qi−j Dj − Ri Dµ−1 , i = γ + µ, γ + µ − 1, · · · , 2, 1, 0,
j=0

where
Ni = 0, i > l,
Qi = 0, i > γ or i < 0,
Ri = 0, i > µ − 1.

Then we have
 2.2 Polynomial Matrices 21

Qi = 0, i > l − µ; (2.17)
 
µ−1
X
Qi = Ni+µ − Qi+µ−j Dj  Dµ−1 , i = l − µ, l − µ − 1, . . . , 1, 0; (2.18)
j=0
i
X
Ri = N i − Qi−j Dj , i = µ − 1, µ − 2, · · · , 1, 0. (2.19)
j=0

It is interesting to note that the degree of Q(s) is l − µ, i.e., max{vi − µi :

i = 1, 2, · · · , m}. If the column degrees of N(s) are strictly less than that of
D(s), i.e., νi < µi for all i, then the quotient Q(s) is zero and R(s) = N (s).
This is because N (s)D(s)−1 is itself strictly proper in this case.

Example 2.2.3. To illustrate the algorithm, consider

 2 
s +3 1  
s + 1 4s − 5
N (s) =  2 −4s2 + s  , D(s) = .
2 s2 − 3s + 2
s + 1 s2 − 4

The matrix D(s) is column-reduced with column degrees µ1 = 1, µ2 = 2, so

µ = max{µ1 , µ2 } = 2. Write D(s) as
       
10 1 4 −1 0 −5 −2 s 0
D(s) = + s + s
01 2 −3 0 2 0 s2
:= [D2 + D1 s−1 + D0 s−2 ]Sc (s).
For N (s), ν1 = ν2 = 2, then l = µ + max{ν1 − µ1 , ν2 − µ2 } = 2 + 1 = 3, and
N (s) is expressed as
        
 10 0 0 30 0 1  
s 0
N (s) =  0 0  s +  0 −4  +  2 1  s−1 +  0 0  s−2
  0 s2
00 1 1 10 0 −4
:= [N3 s + N2 + N1 s−1 + N0 s−2 ]Sc (s).
Since l − µ = 1, Q(s) can be expressed as

Q(s) = Q1 s + Q0 .

From (2.18), we have

 
1 0
Q1 = N3 D2−1 = N3 =  0 0,
0 0

 
  −1 −4
Q0 = N2 − Q1 D1 D2−1 =  0 −4  .
1 1
22 2. Representations of Linear Dynamic Systems

According to (2.15), R(s) should be

R(s) = [R1 s−1 + R0 s−2 ]Sc (s).

It follows from (2.19) that

 
12 −3
R1 = N1 − (Q1 D0 + Q0 D1 ) =  10 −11  ,
−2 −1


0 4
R0 = N0 − Q0 D0 =  0 −8  .
0 −1
Finally, we obtain
 
s − 1 −4
Q(s) = Q1 s + Q0 =  0 −4  ,
1 1
 
12 −3s + 4
R(s) = (R1 s−1 + R0 s−2 )Sc (s) =  10 −11s + 8  . ♦
−2 −s − 1

2.2.3 Stability of Polynomial Matrices

Consider an m × m nonsingular polynomial matrix P (s). P (s) is said to be

stable if det(P (s)) has all roots in C− , the open left-half of the complex
plane. For a nonsingular polynomial matrix P (s), it has been shown before
that there is a unimodular polynomial matrix U (s) such that U (s)P (s) is
row-reduced, i.e., there are integers ν1 , ν2 , · · · , νm such that for some Sr (s) =
diag{sν1 , sν2 , · · · , sνm }, the limit lims→∞ Sr−1 (s)U (s)P (s) = Phr exists and is
nonsingular. Without loss of generality, Phr can be assumed to be an identity
matrix. Therefore, we let the nonsingular polynomial matrix P (s) have the
form
P (s) = Sr (s){Im + Pν−1 s−1 + Pν−2 s−2 + · · · + P1 s−ν+1 + P0 s−ν }, (2.20)

where ν = max{ν1 , ν2 , · · · , νm }. Also, let

Dr (s) = diag (sν−ν1 , sν−ν2 , · · · , sν−νm ).

Then, we have
P̄ := Dr (s)P (s) = Im sν + Pν−1 sν−1 + Pν−2 sν−2 + · · · + P1 s + P0 , (2.21)

det(P̄ (s)) = det(Dr (s)) · det(P (s)) = s∆ν · det(P (s)), (2.22)
 2.2 Polynomial Matrices 23

m
X
where ∆ν = (ν − νi ). Introduce the constant matrix A as follows:
i=1
 
0 0 −P0
 Im −P1 
 
A= . ..  . (2.23)
 .. . 
0 Im −Pν−1

It is well known that

det(P̄ (s)) = det(sI − A). (2.24)

If P (0) is singular, det P (0) is zero, and P (s) has a root at the origin so
that P (s) is unstable. On the other hand, if P (0) is nonsingular, det P (0)
is non-zero so that det P (s) has no roots at the origin. Then, by (2.22), the
non-zero roots of det P (s) will be the same as those of det P (s). We have
actually established the following theorem.

Theorem 2.2.1. Suppose nonsingularity of P (0). The polynomial matrix

P (s) in (2.20) is stable if and only if all the eigenvalues of A in (2.23) are
in C− .

It is believed that the theorem is useful from the computation point of

view because there is a function available in Matlab to compute all eigenvalues
of a matrix.

Example 2.2.4. Let

 2 
s + 4s + 4 s + 2
P (s) = ,
1 s+2

which has nonsingular P (0). It is rewritten as

 2      
s 0 4 1 −1 4 2 −2
P (s) = I2 + s + s .
0 s 12 00

Then, we form
 
0 0 −4 −2
0 0 0 0 
A= 
 1 0 −4 −1  ,
0 1 −1 −2

and obtain its non-zero eigenvalues as −1, −2, −3, which are all in C− . By
Theorem 2.2.1, P (s) is thus stable. ♦
24 2. Representations of Linear Dynamic Systems

2.2.4 Hermite and Smith Forms

Two polynomial matrices P1 (s), P2 (s) ∈ Rp×m [s] are called equivalent (in C)
if there exist unimodular matrices UL (s) ∈ Rp×p [s], UR (s) ∈ Rm×m [s] such
that

UL (s)P1 (s)UR (s) = P2 (s).

With postmultiplying and premultiplying polynomial matrices by appropri-

ate unimodular matrices, it is possible to reduce them to the so called stan-
dard forms. The advantage of a standard form is that it reveals the basic
structure of the matrix and serves as a typical representation of a whole class
of equivalent matrices. Two commonly used standard forms are the Hermite
form (similar to a triangular matrix) if only row or column operations are
allowed; and the Smith form (similar to a diagonal matrix) if both row and
column operations are permitted.
Theorem 2.2.2. (Hermite row form) Let P (s) ∈ Rp×m [s] have rank r.
Then, there exists unimodular matrix U (s) ∈ Rp×p [s] (obtained from elemen-
tary row operations) such that U (s)P (s) = H(s) is in the Hermite row form
which is upper quasi-triangular and defined as follows:
(i) If p > r, the last p − r rows are identically zero;
(ii) In column j, 1 ≤ j ≤ r, the diagonal element is monic (i.e., having
the leading coefficient 1), any element below it is zero, and any element
above it has lower degree than it; and
(iii) If m > r, no particular statements can be made about the elements in
the last m − r columns and first r rows.
Proof. Choose among nonzero elements of the first column of P (s) a polyno-
mial of the lowest degree and bring it to position (1,1) by suitable interchange
of the rows. We denote it by p̄11 (s). Then divide other elements p̄i1 (s) in the
column by p̄11 (s), i.e.

p̄i1 (s) = p̄11 (s)qi1 (s) + ri1 (s), i = 2, . . . , m. (2.25)

Subtracting the i − th row by qi1 times 1st row, we actually replace p̄i1 (s) by
ri1 (s). Among nonzero ri1 (s), choose one with the lowest degree and bring
it to position (1,1) by suitable interchange of the rows. Continuing this pro-
cedure until all elements in the first column are zero except one at position
(1,1). Now consider the second column of the resulting matrix and, temporar-
ily ignoring the first row, repeat the above procedure until all the elements
below the (2,2) element are zero. If the (1,2) element does not have lower
degree than the (2,2) element, the division algorithm and an elementary row
operation can be used to replace the (1,2) element by a polynomial of lower
degree than the diagonal or (2,2) element. Continuing this procedure with
the third column, fourth column, and so on finally gives the desired Hermite
row form. 
 2.2 Polynomial Matrices 25

Remark 2.2.1. By interchanging the roles of rows and columns, one can ob-
tain a similar Hermite column form, the details of which we shall not spell
out.
Example 2.2.5. We proceed as follows:
   
s 3s + 1 1 −(s2 + 2s − 1)
 −1 s2 + s − 2  −r−→
3 ↔r1
 −1 s2 + s − 2 
2
−1 s + 2s − 1 s 3s + 1

   
r1 →r2 1 −(s2 + 2s − 1) 1 −(s2 + 2s − 1)
−−
−sr− −→
1 →r3
0 −(s + 1)
−r2 →r2
 −→  0 (s + 1) 
3 2 2
0 s + 2s + 2s + 1 0 (s + 1)(s + s + 1)

 
(s+1)r2 →r1 1 2
−−−2−−−−−−−→ 
−(s +s+1)r2 →r3 0 s + 1  := H(s)
0 0

which is in the Hermite row form. The corresponding unimodular matrix is

obtained from the above operations (take note of the sequence) as
    
1 (s + 1) 0 1 0 0 1 00 0 0 −1
0 1 0   0 −1 0   1 1 0   0 1 0 
2
0 −(s + s + 1) 1 0 0 1 −s 0 1 10 0

 
0 −(s + 1) s
= 0 −1 1  := U (s). ♦
2 2
1 (s + s + 1) −(s + 1)

Theorem 2.2.3. (Smith form) Let P (s) be a polynomial matrix of rank r,

then there are unimodular polynomial matrices UL (s) and UR (s) such that
 
λ1 (s) 0
 λ2 (s) 
 
 . .. 
UL (s)P (s)UR (s) =   := Λ(s)
 
 λr (s) 
0 0

is the so-called Smith form of P(s) in which each λi (s), i = 1, 2, · · · , r, is a

monic polynomial satisfying the divisibility property:

λi (s)|λi+1 (s), i = 1, 2, · · · , r − 1,

i.e., λi (s) divides λi+1 (s) without remainder. λi (s) are called invariant poly-
nomials of P(s). Moreover, define the determinant divisors of P (s) by ∆i (s)=
26 2. Representations of Linear Dynamic Systems

the monic greatest common divisor (GCD) of all i × i minors of P (s),

i = 1, 2, · · · , r, with ∆0 (s) = 1 by convention. Then, one can identify the
invariant polynomials λi (s) by
∆i (s)
λi (s) = (s), i = 1, 2, · · · , r.
∆i−1

Proof. Bring the least degree element of P (s) to the (1,1) position. As in
construction of the Hermite forms, by elementary row and column operations,
make all other elements in the first row and column zero. If this (1,1) element
does not divide every other element in the matrix, using the polynomial
division algorithm and row and column interchanges, we can bring a lower-
degree element to the (1,1) position, and then repeat the above steps to zero
all other elements in the first row and column. Then we reach
 
λ1 (s) 0 · · · 0
 0 
 
U 1 P V1 =  . ,
 .. P1 (s) 
0
where λ1 divides every element of P1 . Now repeat the above procedure on
P1 . Proceeding in this way gives the Smith form Λ(s). 
Example 2.2.6. Continue Example 2.2.5. With the additional elementary col-
umn operation corresponding to
 
1 −2
:= UR (s),
0 1
we get the Smith form:
 
1 0
H(s)UR (s) =  0 s + 1  := Λ(s). ♦
0 0
Example 2.2.7. Consider a polynomial matrix,
 
1 −1
P (s) =  s2 + s − 4 2s2 − s − 8  .
s2 − 4 2s2 − 8
One proceeds,
∆0 (s) = 1,
∆1 (s) = GCD{1, −1, s2 − s − 4, 2s2 − s − 8, s2 − 4, 2s2 − 8} = 1,
ü ü ü ü
ü 1 −1 ü ü 1 −1 üü
∆2 (s) = GCD{üü 2 ü , ü ,
s + s − 4 2s2 − s − 8 ü ü s2 − 4 2s2 − 8 ü
ü 2 ü
ü s + s − 4 2s2 − s − 8 ü
ü 2 ü}
ü s −4 2s2 − 8 ü
= (s + 2)(s − 2).
 2.2 Polynomial Matrices 27

It follows that
∆1 ∆2
λ1 (s) = = 1, λ2 (s) = = (s + 2)(s − 2).
∆0 ∆1
Thus, one obtains the Smith form:
 
1 0
Λ(s) =  0 (s + 2)(s − 2)  . ♦
0 0

2.2.5 Common Divisor and Coprimeness

Definition 2.2.3. Let P (s), L(s), and R(s) be polynomial matrices such that

P (s) = L(s)R(s).

Then R(s) (resp. L(s)) is called a right (resp. left) divisor of P (s), and P (s)
a left (resp. right) multiple of R(s) (resp. L(s)).

For instance, the polynomial matrix:

 
s(s + 1)(s + 2)
P (s) =  ,
s(s + 1)(s + 3)

has R1 (s) = 1, R2 (s) = s, R3 (s) = s + 1, and R4 (s) = s(s + 1) as its right

divisors. One notes that R4 (s) = s(s+1) is the monic greatest common factor
of two scalar polynomials in P (s).
If P (s) is nonsingular, then all its square divisions (right and left) are
nonsingular, too. For example, the matrix:
 
1 s(s + 1)(s + 2)
P (s) =  ,
0 s(s + 1)(s + 3)

is obviously nonsingular. Its right divisors include

   
10 10
R1 (s) = , R2 = ,
01 0s

   
1 0 1 0
R3 (s) = , R4 = ,
0 s+1 0 s(s + 1)

which are all square and nonsingular.

Definition 2.2.4. Let two polynomial matrices N (s) and D(s) have the
same number of columns.
28 2. Representations of Linear Dynamic Systems

(i) A square polynomial matrix R(s) is called a common right divisor of

N (s) and D(s) if R(s) is a right divisor of both, i.e., there are polynomial
matrices Ñ (s) and D̃(s) such that

N (s) = Ñ (s)R(s),

D(s) = D̃(s)R(s);

(ii) R(s) is called a greatest common right divisor (GCRD) of N (s) and D(s)
if R(s) is a left multiple of any other common right divisor R̄(s) of N (s)
and D(s), i.e. there is a polynomial matrix L(s) such that

R(s) = L(s)R̄(s);

(iii) N (s) and D(s) are called right coprime if all common right divisions of
N (s) and D(s) are unimodular.

Analogously, we can define greatest common left divisors (GCLD) and left
coprimeness of two polynomial matrices having the same number of rows.
Greatest common divisors can be found by elementary operations on a com-
posite matrix formed by N (s) and D(s), as indicated by the following theo-
rem.

Theorem 2.2.4. Let N (s) ∈ Rp×m [s] and D(s) ∈ Rm×m [s].
(i) If a unimodular matrix U (s) ∈ R(p+m)×(p×m) [s] and a square polyno-
mial matrices R(s) ∈ Rm×m[s] are such that
   
D(s) R(s)
U (s) = . (2.26)
N (s) 0

Then, R(s) is a GCRD of N (s) and D(s).

(ii) Further, if there holds

rank [DT (s) N T (s)]T = m, (2.27)

then all GCRDs of N (s) and D(s) must be nonsingular and can differ
only by unimodular (left) divisors. In particular, if a GCRD is unimod-
ular, then all GCRDs must be unimodular.

Proof. (i) Partition U (s) in the form:

 
U11 (s) U12 (s)
U (s) = , (2.28)
U21 (s) U22 (s)

where U11 (s) ∈ Rm×m [s] and U22 (s) ∈ Rp×p [s]. Then the polynomial matrix
U −1 (s) can be partitioned similarly as
 2.2 Polynomial Matrices 29

 −1  
U11 (s) U12 (s) V11 (s) V12 (s)
= .
U21 (s) U22 (s) V21 (s) V22 (s)

It follows that
    
D(s) V11 (s) V12 (s) R(s)
= ,
N (s) V21 (s) V22 (s) 0

or equivalently

D(s) = V11 (s)R(s), N (s) = V21 (s)R(s),

indicating that R(s) is a common right divisor of D(s) and N (s). Notice next
from (2.26) that

R(s) = U11 (s)D(s) + U12 (s)N (s).

Now, if R1 (s) is another common right divisor,

D(s) = D1 (s)R1 (s), N (s) = N1 (s)R1 (s),

then one has

R(s) = [U11 (s)D1 (s) + U12 (s)N1 (s)]R1 (s),

so that R1 (s) is a right divisor of R(s), and thus R(s) is a GCRD.

(ii) Since elementary operations do not change rank, [RT , 0]T has full
rank or R is nonsingular if (2.27) holds true. If R̄ is any other GCRD, then
R(s) and R̄(s) are related by definition as

R(s) = L1 (s)R̄(s), R̄(s) = L2 (s)R(s),

for polynomial Li . Thus, R̄(s) is nonsingular. And we can write

R(s) = L1 (s)L2 (s)R(s),

which implies det(L1 ) det(L2 ) = 1 and both L1 and L2 are unimodular. 

Notice that the reduction of [DT (s) N T (s)]T to [RT (s) 0]T can be
achieved by elementary row operations such that the last p rows from the
bottom on the right-hand side become zero. In particular, we could use the
procedure for getting the Hermite row form, but this is more than necessary.

Example 2.2.8. Find a GCRD of

 
s 3s + 1
D(s) = , N (s) = [−1 s2 + 2s − 1].
−1 s2 + s − 2

It follows from Example 2.2.5 that there is a unimodular matrix U (s) such
that
30 2. Representations of Linear Dynamic Systems

 
  1 2
D(s)  0 s + 1
U (s)  · · ·  =  
··· ··· .
N (s)
0 0

Then, one knows from Theorem 2.2.4 that

 
1 2
R(s) =
0 s+1

is a GCRD of the given D(s) and N (s). ♦

Theorem 2.2.5. Suppose full rank of [DT (s) N T (s)]T . N (s) and D(s) are
right coprime if and only if one of the following holds true:
(i) There exist polynomial matrices X(s) and Y (s) such that

X(s)N (s) + Y (s)D(s) = I; (2.29)

(ii) [DT (s) N T (s)]T has full rank for every s ∈ C.

Proof. (i) Based on the Theorem 2.2.4, it is always possible to write a GCRD
R(s) of N (s) and D(s) as R(s) = X̂(s)N (s) + Ŷ (s)D(s) for polynomial X̂
and Ŷ . Moreover, if N (s) and D(s) are coprime, R(s) must be unimodular
so that
I = R−1 (s)[X̂(s)N (s) + Ŷ (s)D(s)]
= X(s)N (s) + Y (s)D(s),

where X(s) := R−1 (s)X̂(s) and Y (s) := R−1 (s)Ŷ (s) are both polynomial.
Conversely, suppose that there exist two matrices X(s) and Y (s) satisfying
(2.29). Let R(s) be a GCRD of N (s) and D(s), i.e.

N (s) = N̂ (s)R(s), D(s) = D̂(s)R(s).

Equation (2.29) then becomes

h i
I = X(s)N̂ (s) + Y (s)D̂(s) R(s),

yielding

R−1 (s) = X(s)N̂ (s) + Y (s)D̂(s),

a polynomial matrix. This entails that R(s) is a unimodular matrix so that

N (s) and D(s) are right coprime.
(ii) From (2.26) we see that [DT (s) N T (s)]T has full rank for all s if
and only if R(s) is unimodular, i.e., if and only if all GCRDs of N (s) and
D(s) are unimodular, which by definition means that N (s) and D(s) are right
coprime. 
 2.3 Rational Matrices and Polynomial Matrix Fractions 31

Example 2.2.9. Check right coprimeness of

 2 
s + 2s s + 3  
D(s) = 2 , N (s) = s 1 .
2s − s 3s − 2
One readily sees that
 
  0 3
D(s)
=  0 −2 
N (s) s=0
0 1

is of rank defect, and then D(s) and N (s) are not coprime. One may write
 2   
  s + 2s s + 3 s+2 s+3  
D(s)  2    s0
= 2s − s 3s − 2 = 2s − 1 3s − 2
N (s) 01
s 1 1 1
 
D̃(s)
:= R(s).
Ñ (s)
and find that the right divisor R(s) is not unimodular as detR(s) = s =
6 const,
which is in agreement with non-coprimeness of D(s) and N (s). One notes that
h iT
the first and third rows of D̃T (s) Ñ T (s) form a nonsingular matrix for
any s ∈ C. Thus, D̃(s) and Ñ (s) are right coprime and R(s) is actually a
GCRD of them. ♦

2.3 Rational Matrices and Polynomial Matrix Fractions

A rational function matrix G(s) is a matrix whose elements are rational func-
tions. We are interested to generalize polynomial fractions of scalar rational
functions to the matrix case.
Definition 2.3.1. A right polynomial matrix fraction (PMF) for the p × m
rational function matrix G(s) is an expression of the form
G(s) = N (s)D−1 (s),
where N (s) is a p× m polynomial matrix, and D(s) is an m× m nonsingular
polynomial matrix. Such a PMF is further called a right coprime PMF if
N (s) and D(s) are right coprime. The degree of a right PMF is the degree
of the polynomial detD(s). A left polynomial matrix fraction for G(s) is an
expression of the form
−1
G(s) = DL (s)NL (s),
where NL (s) is a p× m polynomial matrix, and DL (s) is a p× p nonsingular
polynomial matrix. Similar definitions apply for the coprimeness and degree
of a left PMF.
32 2. Representations of Linear Dynamic Systems

Of course, this definition is familiar if m = p = 1. In the multi-input and

multi-output case, a simple device can be used to exhibit PMFs for a given
G(s). Suppose that d(s) is a least common denominator of all elements of
G(s). (In fact, any common denominators can be used). Then

Nd (s) = d(s)G(s)

is a p × m polynomial matrix, and we can write either a right or left PMF:

G(s) = Nd (s)[d(s)Im ]−1 = [d(s)Ip ]−1 Nd (s). (2.30)

The degrees of these two PMFs are different in general, and it should not be
surprising that lower-degree PMFs can be found if some effort is invested.
Let dci (s) (resp. dri (s)) be a least common denominator of all elements in
the ith column (resp. ith row) of G(s), then

N (s) := G(s)diag{dc1 , dc2 , · · · , dcm }

and

NL (s) := diag{dr1 , dr2 , · · · , drp }G(s)

are both polynomial. We obtain PMFs as

G(s) = N (s) · diag −1 {dc1 , dc2 , · · · , dcm }
= diag −1 {dr1 , dr2 , · · · , drp } · NL (s).
Example 2.3.1. Consider a rational matrix
" 1 1
#
s+1 s+2
G(s) = 1
.
0 s+3

A least common denominator of it is d(s) = (s + 1)(s + 2)(s + 3) so that

 
(s + 2)(s + 3) (s + 1)(s + 3)
G(s) = [d(s)I]−1
0 (s + 1)(s + 2)
 
−1 (s + 2)(s + 3) (s + 1)(s + 3)
= [d(s)I]
0 (s + 1)(s + 2)
where two PMFs are both of degree 6 but neither of them is coprime. On the
other hand, one can write
  −1
1 (s + 3) (s + 1) 0
G(s) =
0 (s + 2) 0 (s + 2)(s + 3)
 −1  
(s + 1)(s + 2) 0 (s + 2) (s + 1)
= ,
0 (s + 3) 0 1
where the degree of two fractions is now 3 and both of them are actually
coprime. ♦
 2.3 Rational Matrices and Polynomial Matrix Fractions 33

2.3.1 Coprime Polynomial Matrix Fractions

In the single-input and single output case, the issue of common factors in the
numerator and denominator polynomials of G(s) draws attention. The utility
of PMFs begins to emerge from the corresponding concept in the matrix case.
For PMFs, one of the polynomial matrices is always nonsingular, so only
nonsingular common divisors occur. Suppose that G(s) is given by a right
PMF:

G(s) = N (s)D−1 (s),

and that R(s) is a common right divisor of N (s) and D(s). Then
e (s) = N (s)R−1 (s)
N and e
D(s) = D(s)R−1 (s) (2.31)
are polynomial matrices, and they provide another right PMF for G(s) since
e (s)D
N e −1 (s) = N (s)R−1 (s)R(s)D−1 (s) = G(s).

The degree of this new PMF is no greater than the degree of the original,
since
e
deg [det D(s)] = deg [det D(s)] + deg [det R(s)].

Of course, the largest degree reduction occurs if R(s) is a greatest common

right divisor, and no reduction occurs if N (s) and D(s) are right coprime.
This discussion indicates that extracting common right divisors of a right
PMF is a generalization of the process of cancelling common factions in a
scalar transfer function.

Theorem 2.3.1. For any p × m rational function matrix G(s), there is

a right coprime PMF; Further, for any two right coprime PMFs, G(s) =
N (s)D−1 (s) = N1 (s)D1−1 (s), there exists a unimodular polynomial matrix
U (s) such that N (s) = N1 (s)U (s) and D(s) = D1 (s)U (s).

Proof. Given a G(s), it is always possible, say by (2.30), to get a right PMF,
G(s) = N (s)D−1 (s). Using Theorem 2.2.4, we can extract a GCRD, R(s), of
N (s) and D(s) from
   
D(s) R(s)
U (s) = . (2.32)
N (s) 0
−1
A right coprime PMF is then obtained as NR (s)DR (s) where

NR (s) = N (s)R−1 (s), DR (s) = D(s)R−1 (s).

Next, for any two right coprime PMFs, G(s) = N (s)D−1 (s) = N1 (s)D1−1 (s),
by Theorem 2.2.5, there exist polynomial matrices X(s), Y (s), A(s), and
B(s) such that
34 2. Representations of Linear Dynamic Systems

X(s)N1 (s) + Y (s)D1 (s) = Im , (2.33)

and

A(s)N (s) + B(s)D(s) = Im . (2.34)

Since N (s)D−1 (s) = N1 (s)D1−1 (s), we have N1 (s) = N (s)D−1 (s)D1 (s). Sub-
stituting this into (2.33) gives

X(s)N (s)D−1 (s)D1 (s) + Y (s)D1 (s) = Im ,

or

X(s)N (s) + Y (s)D(s) = D1−1 (s)D(s).

A similar calculation using N (s) = N1 (s)D1−1 (s)D(s) in (2.34) gives

A(s)N1 (s) + B(s)D1 (s) = D−1 (s)D1 (s).

One sees that both D1−1 D(s) and D−1 (s)D1 (s) are polynomial matrices, but
they are inverses of each other. Thus, both must be unimodular. Let

U (s) = D1−1 (s)D(s).

Then, we have

N (s) = N1 (s)U (s), D(s) = D1 (s)U (s).

The proof is complete. 

Surprisingly, the extraction procedure in (2.32) gives us more than a right

coprime PMF. Partition U (s) as
 
YR XR
U (s) = ,
−NL DL

and U −1 (s) as
 
DR −XL
U −1 (s) = .
NR YL

Then, there holds

    
−1 YR XR DR −XL I0
UU = = . (2.35)
−NL DL NR YL 0I

Theorem 2.3.2. Let N (s)D−1 (s) is a right PMF of G(s) with the operations
−1
leading to (2.32) performed. Then NR DR is a right coprime PMF of G(s),
−1
and DL NL is a left coprime PMF of G(s).
 2.3 Rational Matrices and Polynomial Matrix Fractions 35

Proof. Premultiplying (2.32) by U −1 yields

      
D DR −XL R DR R
= = ,
N NR YL 0 NR R

so that D = DR R with DR nonsingular since so is D. Furthermore, one sees

−1 −1
NR D R = NR RR−1 DR = (NR R)(DR R)−1 = N D−1 = G.

The (1, 1) block of (2.35) gives

YR DR + XR NR = I,
−1
where YR and XR are both polynomial. Thus NR DR is a right coprime MFD
of G(s).
The (2,1) block of (2.35) gives
D L NR = NL D R . (2.36)
We now show nonsingularity of DL by contradiction. Assume that there is
a non-zero polynomial vector α(s) such that αT DL = 0. Then by (2.36),
one has 0 = αT DL NR = αT NL DR . This implies αT NL = 0 since DR is
nonsingular. It then follows that
αT [NL XL + DL YL ] = αT NL XL + αT DL YL = 0.
But the (2,2) block of (2.35):
NL XL + DL YL = I, (2.37)
shows

αT [NL XL + DL YL ] = αT I = αT 6= 0.
−1 −1 −1
From (2.36), DL NL = NR D R = G(s), so that DL NL is a left PMF of
G(s), and it is coprime in view of (2.37). 

Example 2.3.2. Consider G(s) = N (s)D−1 (s) where

 
s (3s + 1)  
D(s) = , N (s) = −1 (s2 + 2s − 1) .
−1 (s2 + s − 2)

It follows from Example 2.2.5 that U which makes U [DT N T ]T = [RT 0]T
is
 
0 −(s + 1) s
U = 0 −1 1 ,
2 2
1 (s + s + 1) −(s + 1)
−1
from which we get a left coprime PMF, DL (s)NL (s), of G(s) where
36 2. Representations of Linear Dynamic Systems

DL (s) = −(s2 + 1), NL (s) = −[1 (s2 + s + 1)].

Further, one calculates

 
s 1 ∗
U −1 =  −1 s ∗  ,
−1 s + 1 ∗

so that a right coprime PMF is obtained as

 −1
−1
  s 1
NR D R = −1 (s + 1) . ♦
−1 s
−1
Theorem 2.3.3. (Generalized Bozout Identity) Let NR DR be a right
coprime PMF. Then there are six polynomial matrices XR , YR , NL , DL , XL
and YL such that
    
YR XR DR −XL I0
= . (2.38)
−NL DL NR YL 0I

Proof. Since NR and DR are right coprime, there are polynomials XR and
YR such that YR DR + XR NR = I. By Theorem 2.3.2, we can have a left
−1 −1
coprime polynomial matrices NL and DL such that DL NL = NR D R . The
left coprimeness of NL and DL implies the existence of polynomials X̃L and
eL = I. Writing these relations in matrix form gives
ỸL such that DL YeL +NL X
 " #  
YR XR DR −X eL IQ
= ,
−NL DL NR YeL 0 I

eL + XR YeL . Post-multiplying both sides by

where Q = −YR X
 −1  
IQ I −Q
=
0 I 0 I

produces (2.38), where XL = DR Q + X̃L and YL = YeL − NR Q. 

2.3.2 Smith-McMillan Form

Write a rational matrix G(s) as G(s) = Nd(s) (s)

, where d(s) is the monic
least common denominator of all elements of G(s). Apply Theorem 2.2.3
to the polynomial matrix, N (s) = G(s)d(s), to get its unique Smith form
Λ(s) = diag[λ1 (s), λ2 , · · · , λr , 0] with N (s) = UL (s)Λ(s)UR (s) holding for
some unimodular matrices UL (s) and UR (s). Using G(s) = UL (s) Λ(s) d(s) UR (s)
and cancelling all common factors for each pair of λi and d, we will get the
Smith-McMillan form.
 2.3 Rational Matrices and Polynomial Matrix Fractions 37

Theorem 2.3.4. (Smith-McMillan form) Let a rational G ∈ Rp×m (s)

have rank r. Then there exist unimodular matrices UL ∈ Rp×p [s] and UR ∈
Rm×m [s] such that

G = UL M U R ,

and
 
1 /ψ1 0
 2 /ψ2 
 
 .. 
M = .  ∈ Rp×m (s) (2.39)
 
 r /ψr 
0 0
is the so called Smith-McMillan form of G(s), where (i , ψi ) are pairs of
monic polynomials copime to each other, satisfying the following divisibility
properties

ψi+1 |ψi , i = 1, 2, · · · , r − 1,

i |i+1 , i = 1, 2, · · · , r − 1,

where ψ1 = d is the monic least common divisor of all the elements of G.

Example 2.3.3. Let

 1 −1 
s2 +3s+2 s2 +3s+2
 2 
 +s−4 2s2 −s−8 
G(s) =  ss2 +3s+2 .
s2 +3s+2 

s−2 2s−4
s+1 s+1

It is written as
1
G(s) = P (s),
(s + 1)(s + 2)

where P (s) is given in Example 2.2.7. So

 1

(s+1)(s+2)
0
Λ(s) 

s−2 
= 0  := M (s)
s+1 
d(s) 
0 0
is the Smith-McMillan form of G(s). ♦
38 2. Representations of Linear Dynamic Systems

Given the Smith-McMillan form of G(s), it is straightforward to arrive at

coprime factorization of G. Define
 
diag(1 , 2 , · · · , r ) 0
W = ∈ Rp×m [s],
0 0
 
diag(ψ1 , ψ2 , · · · , ψr ) 0
ΨR = ,
0 Im−r
 
diag(ψ1 , ψ2 , · · · , ψr ) 0
ΨL = .
0 Ip−r
G could be expressed in two different ways as follows
G = UL M UR = (UL W )(UR−1 ΨR )−1 := NR DR
−1
, (2.40)
= (ΨL UL−1 )−1 (W UR ) := −1
DL NL . (2.41)
Note that even though the non-zero polynomial elements of Ψr and Ψl are
the same, ΨL is a p × p matrix and ΨR is a m × m matrix.
We now show that the polynomial matrix pair (NR , DR ) is right coprime.
Write
   −1   −1  
DR U R ΨR UR 0 ΨR
= = .
NR UL W 0 UL W
   
DR ΨR
and have the same rank because the transformation matrix
NR W  
ΨR
relating them is unimodular. Consider the rank of . The first r columns
 W

ΨR
of are linearly independent and the individual columns do not vanish
W
for any s ∈ C. This follows from the fact that the only non-zero elements
of the ith column, namely ψi and i , do not vanish simultaneously for any
s ∈ C as they are coprime to each other. The remaining (m − r) columns
have only one non-zero element, namely unity
 for each column and they are
ΨR
linearly independent. Hence the rank of is equal to m for all s ∈ C.
W
−1
The matrix pair (UL W, UR ΨR ) is therefore right coprime. Similarly it can
be shown that (ΨL UL−1 , W UR ) is left coprime.
Using the above fact and Theorem 2.3.4, we establish the following theo-
rem.

Theorem 2.3.5. Let G ∈ Rp×m (s).

−1 −1
(i) NR D R and DL NL in (2.40) and (2.41) are right and left coprime
PMFs of G(s), respectively;
(ii) All the numerator matrices, whether they belong to the right or left co-
prime PMFs of G, are equivalent and have the same Smith form as
W;
 2.3 Rational Matrices and Polynomial Matrix Fractions 39

(iii) The denominator matrices of any (right or left) coprime PMFs of G

have the same non-unity invariant polynomials. Hence, in particular,
their determinants have the same degree and the same roots. Moreover,
if G(s) is square, then these denominators are equivalent.

2.3.3 Poles and Zeros

Definition 2.3.2. (poles and zeroes) Let G(s) be a rational matrix with
the Smith-McMillan form M (s) in (2.39), define the pole (or characteristic)
polynomial and zero polynomial of G(s), respectively, as
p(s) = ψ1 (s)ψ2 (s) · · · ψr (s),
z(s) = ε1 (s)ε2 (s) · · · εr (s).
The degree of p(s) is called the McMillan degree of G(s).

Example 2.3.4. The pole and zero polynomials of G(s) in Example 2.3.3 are
p(s) = (s + 1)2 (s + 2),
z(s) = (s − 2).
Hence G(s) has poles {−1, −1, −2} and a zero {2}. ♦

The following is a straightforward corollary of Theorem 2.3.5.

Corollary 2.3.1. Let G(s) ∈ Rp×m (s) have rank r, and N (s) and D(s) be
the numerator and denominator of any coprime (right or left) PMF of G(s),
respectively. Then,
(i) ρ is a pole of G(s) if and only if detD(ρ) = 0;
(ii) ζ is a zero of G(s) if and only if rank N (ζ) < r.

In view of the above corollary, one can find poles and zeros of G(s) from
any coprime PMF of it (one may not necessarily get its Smith-McMillan
form). It is easily shown that ρ is a pole of G(s) if and only if it is a pole
of some element of G(s). But one can hardly tell its multiplicity in p(s)
from individual elements. The case is worse for zeros: a zero of elements in
G(s) has no relationship with zeros of G(s). Another interesting property of
multivariable poles and zeros is that a multivariable system can have a pole
and a zero at the same place, but they do not cancel each other. For instance,
 1 
s+1 0
G= ,
s+1
0 s

has a pole and a zero both at s = −1.

40 2. Representations of Linear Dynamic Systems

Definition 2.3.3. For a rational matrix G(s), if there are polynomial ma-
trices NL , D and NR such that NR D−1 NL = G, then NL D−1 NR is called a
right-left PMF of G. Furthermore, if NR and D are right coprime and D and
NL are left coprime, NR D−1 NL is called a right-left coprime PMF. Obvi-
ously, for NL = I (resp. NR = I), NR D−1 NL reduces to normal right (resp.
left) PMF NR D−1 (resp.NL D−1 ).
Theorem 2.3.6. For a right-left coprime PMF, G = NR D−1 NL , det D is
the pole polynomial of G(s) modulo a nonzero constant.
−1
Proof. Let N L D be a right coprime PMF of D−1 NL , so that D−1 NL =
−1
N L D and detD = detD modulo a nonzero real number. Noting that
(NR N L )(D)−1 = G and Theorem 2.3.5, the proof would be completed if we
show that (NR N L , D) is a right coprime.
Since both pairs, (NR , D) and (N L , D), are coprime, respectively, there
are polynomials X, Y , X and Y such that
XD + Y NR = I, (2.42)

X D + Y N L = I. (2.43)
By postmultiplying (2.42) by D−1 NL , we get successively
XNL + Y NR D−1 NL = D−1 NL ,

−1 −1
XNL + Y NR N L D = N LD ,

XNL D + Y NR N L = N L .
Substituting this into (2.43) yields
(X + Y XNL )D + (Y Y )NR N L = I,
indicating coprimeness of NR N L and D. 
Example 2.3.5. Consider a general state-space model:
ẋ(t) = Ax(t) + Bu(t),
y(t) = Cx(t).
Its transfer function matrix is given by
G(s) = C(sI − A)−1 B.
So a state-space realization of the system naturally gives rise to a right-left
PMF of the transfer matrix. Furthermore, if the model is both observable
and controllable or minimal, then by Theorem Theorem 2.1a, (C, sI − A) is a
right coprime, and (sI − A, B) is left coprime, and C(sI − A)−1 B is actually
a right-left coprime fraction of G(s). It follows form Theorem Theorem 2.13a
that det(sI − A) coincides with the pole polynomial of G(s). ♦
 2.3 Rational Matrices and Polynomial Matrix Fractions 41

2.3.4 Proper Rational Matrices and Realization

Recall from Section 2.1 that a rational transfer function matrix G(s) is said
to be proper if lims→∞ G(s) < ∞ and strictly proper if lims→∞ G(s) = 0. In
the scalar case, a transfer function is proper if the degree of the numerator
polynomial is less than or equal to the degree of the denominator polynomial.
The situation is not so simple in the matrix case, as we shall now explore.

Lemma 2.3.1. Let G(s) ∈ Rp×m (s) be proper (resp. strictly proper) and
have a right PMF, N (s)D−1 (s). Then, every column of N (s) has degree less
than or equal to (resp. strictly less than) that of the corresponding column of
D(s).

Proof. We have N (s) = G(s)D(s), or equivalently, elementwise,

m
X
nij (s) = gik dkj , i = 1, 2, · · · , p.
k=1

Let now µj = ∂cj D(s), the j-th column degree of D(s), and evaluate
m
X
lim nij (s)s−µj = ( lim gik (s))( lim dkj (s)s−µj ),
s→∞ s→∞ s→∞
k=1

which is finite (resp. zero) if gik is proper (resp. strictly proper). Hence we
have ∂cj N (s) ≤ ∂cj D(s) (resp. ∂cj N (s) < ∂cj D(s)). 

However, somewhat surprisingly, the converse of the result is not always

true. For example, let
 2 
s s2
N (s) = [s 1], D(s) = 2 2 .
s s +1

The degrees of the columns of N (s) are less than those of the corresponding
columns of D(s). But
h 2 i
N (s)D−1 (s) = s −s+1
s (1 − s)

is not proper.
To obtain necessary and sufficient conditions for the properness of N (s)×
D−1 (s), suppose that D(s) is column-reduced with its highest column degree
coefficient matrix Dhc being nonsingular. Let Sc (s) = diag[sµ1 , sµ2 , ..., sµm ] .
Then one can write

N (s)D−1 (s) = [N (s)Sc−1 (s)][D(s)Sc−1 (s)]−1 ,

and take the limit

42 2. Representations of Linear Dynamic Systems

lim N (s)D−1 (s) = lim [N (s)Sc−1 (s)] lim [D(s)Sc (s)−1 ]

s→∞ s→∞ s→∞
−1
= Nhc Dhc ,
where Nhc is finite (resp. zero) if ∂cj N (s) ≤ ∂cj D(s) (resp. ∂cj N (s) <
∂cj D(s)). Hence, we have the following theorem.

Theorem 2.3.7. If D(s) is column-reduced, then G(s) = N (s)D−1 (s) is

strictly proper (resp. proper) if and only if each column of N (s) has degree
less than (resp. less than or equal to) the degree of the corresponding column
of D(s).

If G(s) ∈ Rp×m (s) is proper, we can have a state space realization (A, B,
C, D) in (2.1-2.2). Obviously, one sees D = G(∞) and G(s) − G(∞) = G(s)
will be strictly proper. We thus aims to get A, B and C from a right PMF
N̄ (s)D̄−1 (s) of a strictly proper G(s). It follows from Proposition 2.2.1 that
there is a unimodular matrix U (s) such that D(s) = D̄(s)U (s) is column-
reduced. So, without loss of generality, we assume that N (s)D−1 (s) is a
right PMF of a p × m strictly proper G(s) with D(s) column-reduced and
∂cj N (s) < ∂cj D(s) = µj , j = 1, 2, · · · , m. Let n = Σµj and write D(s) as

D(s) = Dhc Sc (s) + Dlc Φc (s)

for suitable constant matrices Dhc and Dlc , where Sc (s) = diag{sµ1 , sµ2 , · · · ,
sµm } as before,

Φc (s) = block diag{φc1 (s), φc2 (s), · · · , φcm (s)},

φci (s) = [sµi −1 , sµi −2 , · · · , s, 1]T , i = 1, 2, . . . , m.

Similarly, we write

N (s) = Nc Φc (s)

for a suitable constant matrix Nc . Define

Ao = block diag {Ao1 , Ao2 , · · · , Aom },
 
0 0
 .. 
1 . 

Aoj =   ∈ Rµj ×µj , j = 1, 2, · · · , m,
. . 
 .. .. 
0 1 0
Bo = block diag {bo1 , bo2 , · · · , bom },
b0j = [1, 0, · · · , 0]T ∈ Rµj ×1 , j = 1, 2, . . . , m.
We now claim that a controllable realization of G(s) is given by
 2.3 Rational Matrices and Polynomial Matrix Fractions 43

−1
A = Ao − Bo Dhc Dlc , (2.44)
−1
B= Bo Dhc , (2.45)
C = Nc . (2.46)
We write D(s) as
−1
D(s) = Dhc (Sc (s) + Dhc Dlc Φc (s))
:= Dhc Γ (s),
and look at
−1
(sI − A)Φc (s) = (sI − Ao + Bo Dhc Dlc )Φc (s)
−1
= (sI − Ao )Φc (s) + Bo Dhc Dlc Φc (s)
−1
= Bo Sc (s) + Bo Dhc Dlc Φc (s)
= Bo Γ (s),
which can be written as

(sI − A)−1 Bo = Φc (s)Γ −1 (s).

−1
Post-multiplying the above by Dhc gives
−1
(sI − A)−1 Bo Dhc = Φc (s)[Dhc Γ (s)]−1 ,

or

(sI − A)−1 B = Φc (s)D−1 (s).

Pre-multiplying by C = Nc yields

C(sI − A)−1 B = N (s)D−1 (s),

and hence (A, B, C) is a realization of N (s)D−1 (s). Furthermore, (Ao , Bo ) is

controllable, since by inspection, Rank[sI − Ao Bo ] has full row rank for all
s. Write
 
I 0
[sI − A B] = [sI − Ao Bo ] −1 −1 ,
Dhc Dlc Dhc

which is also of full row rank due to nonsingularity of the second matrix on
the right hand side. Thus, (A, B) is controllable.

Theorem 2.3.8. Let N (s)D−1 (s) be a right PMF of a strictly proper ratio-
nal function matrix G(s) with D(s) column-reduced. Then
(i) (A, B, C) in (2.44)-(2.46) is a controllable realization of G(s);
(ii) (A, B, C) is a controllable and observable (or minimal) realization of
G(s) if N (s)D−1 (s) is coprime; and
44 2. Representations of Linear Dynamic Systems

(iii) det(sI − A) = α detD(s) for a nonzero real number α if N (s)D−1 (s) is

coprime. Hence, det(sI − A) and the determinants of the denominators
in any coprime fractions of G(s) have the same set of roots precisely,
and dim A is equal to the McMillan degree of G(s).

The proof of (ii) and (iii) can be found, say, in Chen (1984). With obvious
changes, dual results to Theorems 2.3.7 and 2.3.8 can be established for
properness and realization of left PMFs.

Example 2.3.6. Consider a transfer matrix

 
2s+3 1
(s+1)(s+2) s+2 
G(s) =  .
1 1
s+2 s+2

It is easy to get a right PMF,

  −1
(2s + 3) 1 (s + 1)(s + 2) 0
G(s) = := Ñ (s)D̃−1 (s).
(s + 1) 1 0 (s + 2)
After extracting its GCRD,
 
1 −1
R(s) = ,
0 s+2

One gets
  1 
−1 (s + 1)(s + 2) 0 1 s+2
D̄(s) := D̃(s)R (s) = 1
0 s+2 0 s+2
 
(s + 1)(s + 2) s + 1
=
0 1
and
  1 
−1 2s + 3 1 1 s+2
N̄ (s) := Ñ (s)R (s) = 1
s+1 1 0 s+2
 
2s + 3 2
=
s+1 1
for a right coprime PMF N̄ (s)D̄−1 (s) of G(s).
But D̄(s) is not column reduced, since its highest column degree coefficient
matrix,
 
11
D̄hc = ,
00

is singular. We see that (1,1) element will become zero if the 2nd column is
multiplied by −(s+2) and the result is added to the first column. This means
 2.3 Rational Matrices and Polynomial Matrix Fractions 45

  
(s + 1)(s + 2) s + 1 1 0
D̄(s)U (s) =
0 1 −(s + 2) 1
 
0 s+1
= := D(s),
−(s + 2) 1
which is now column reduced as desired. The corresponding numerator in the
right coprime PMD N (s)D−1 (s) = G(s) is
  
2s + 3 2 1 0
N (s) := N̄ U (s) =
s+1 1 −(s + 2) 1
 
−1 2
= .
−1 1
Using the realization formulas, we have
 µ   
s 1 0 s0
Sc (s) = = ,
0 sµ2 0s

 
10
Φc (s) = = I2 ,
01

      
0 s+1 0 1 s0 0 1
D(s) = = + I ,
−(s + 2) 1 −1 0 0s −2 1 2

 
−1 2
N (s) = I = Nc Φ c ,
−1 1 2

   
00 10
Ao = , Bo = ,
00 01

 
−1 −1 −2 1
A = Ao − Bo Dhc Dlc = −Dhc Dlc = ,
0 −1


−1 −1 0 −1
B= Bo Dhc = Dhc = ,
1 0

 
−1 2
C = Nc = .
−1 1

Besides, (A, B, C) is minimal. ♦

46 2. Representations of Linear Dynamic Systems

2.4 Model Reduction

Reduced-order models are often required for simplifying the design and im-
plementation of control systems. A reduced-order model is usually adequate,
provided it has dynamic characteristics close to that of the original high-order
system in the frequency range which is most relevant for control system de-
sign. Therefore, model reduction has been an active research area in engineer-
ing, especially in model-based prediction, control and optimization, and is a
key step in many designs to be presented in later chapters. For plants with
rational transfer functions, balanced truncation and optimal Hankel norm
approximation are popular. However, the transfer functions which will ap-
pear in the subsequent chapters usually contain time delay and may even
not fit into the form of a rational function plus time delay. In this section,
we present two approaches to model reduction, based on the recursive Least
Squares and step response construction, respectively. A great attention is
paid to the preservation of stability for the reduced-order models. These two
approaches are generally applicable provided that plant transfer function or
frequency response is available.

2.4.1 Recursive Least Square Reduction

Consider a scalar system of possibly complicated dynamics with its transfer

function G(s) (probably non-rational) or frequency response G(jω) available.
The problem at hand is to find a nth-order rational function plus dead time
model:
βn sn + · · · + β1 s + β0
Ĝ(s) = ĝ0 (s)e−Ls = e−Ls , (2.47)
sn + αn−1 sn−1 + · · · + α1 s + α0

such that the approximation error e defined by

M ü
X ü2
ü ü
e= üW (jωi )(G(jωi ) − Ĝ(jωi ))ü , (2.48)
i=1

is minimized, where the interval [ω1 , ωM ] defines the frequency range of

interest and W (jωk ) serves as the weighting. Note that (2.47) contains the
unknown time delay L to be estimated and it makes the problem nonlinear.
The solution to our original model reduction is obtained by minimizing the
approximation error over the possible range of L. This is a one-dimensional
search problem and can be easily solved if an estimation of the range of L
is given. A reasonable search range of L is 0.5 ∼ 2.0 times of L0 which is
the initial estimate of the time delay of G(s), and evaluate 15 ∼ 20 points in
that range to find the optimal estimate L̂. An alternative is to use a Newton-
Raphson type of algorithms to search for the optimal L (Lilja, 1989), which
requires no prior range for L.
 2.4 Model Reduction 47

If the time delay L is known, then the problem becomes to approximate

a modified plant g0 (s) = G(s)eLs with a rational transfer function
βn sn + · · · + β1 s + β0
ĝ0 (s) = ,
sn + αn−1 sn−1 + · · · + α1 s + α0
such that
M
X 2
e0 , |W (jωi )(g0 (jωi ) − ĝ0 (jωi ))| , (2.49)
i=1

is minimized. Equation (2.49) falls into the framework of transfer function

identification in frequency domain. For this identification, a number of meth-
ods are available (Pintelon et al., 1994), and the recursive least square (RLS)
algorithm is simple and effective and is briefly described as follows.
M
X
(k) (k) (k) (k) (k)
e0 , |W̄i {g0 (jωi )[(jωi )n + αn−1 (jωi )n−1 + ... + α1 (jωi ) + α0 ]
i=1
(k) (k)
−[βn(k) (jωi )n + ... + β1 (jωi ) + β0 ]}|2 , (2.50)

where
(k) W (jωi )
W̄i , (k−1) (k−1) (k−1)
(2.51)
(jωi )n + αn−1 (jωi )n−1 + ... + α1 (jωi ) + α0
acts as a weighting function in the standard least squares problem in braces.
Equation (2.50) is re-arranged to yield
M
X
(k) (k) (k)T (k) 2
e0 , |ηi − φi θ | ,
i=1

where
(k) (k)
ηi = −g0 (jωi )(jωi )n W̄i , (2.52)
(k) (k) (k) (k) (k) (k) T
θ = [ αn−1 ... α0 βn ... β1 β0 ] , (2.53)
(k) (k)
φi = [ g0 (jωi )(jωi )n−1 . . . g0 (jωi ) −(jωi )n . . . −(jωi ) −1 ]T W̄i . (2.54)

Then, we have the recursive equation for θ(i) as

θ(k,i) = θ(k,i−1) + K (k,i) ε(k,i) , i = 1, 2, · · · , M, (2.55)
where
(k) (k)T (k,i−1)
ε(k,i) = ηi − φi θ . (2.56)
(k,i) (k) (k)T (k)
K = P (k,i−1) φi (I + φi P (k,i−1) φi )−1 , (2.57)
(k,i) (k)T
P = (I − K (k,i) φi )P (k,i−1) , (2.58)
48 2. Representations of Linear Dynamic Systems

Once the above RLS in (2.55)-(2.56) has been completed, the resultant pa-
(k)
rameter vector θ(k) = θ(k,M ) is used to update W̄i to

(k+1) 1
W̄i = (k) (k) (k)
, (2.59)
(jωi )n + αn−1 (jωi )n−1 + ... + α1 (jωi ) + α0

and (2.55)-(2.56) are repeated to calculate θ(k+1) . On convergence, the resul-

tant parameter vector will form one solution to (2.49).
Different weighting functions are employed in the various methods
(k)
(Pintelon et al., 1994). For simplicity, it is recommended that the W̄i is
chosen as
1
(k−1) (k−1) (k)
.
(jωi )n + αn−1 (jωi )n−1 + ... + α1 (jωi ) + α0

Simulation shows that the most important frequency range for the model
reduction is a decade above and below ωc , where ωc is the unity gain cross-
over frequency of the transfer function G0 (s). Therefore, the frequency range
[ω1 , ωM ] in the optimal fitting problem (2.49) is chosen to span M logarith-
mically equally spaced points between 0.1ωc and 10ωc .
Once a reduced-order model Ĝ(s) is found, the following frequency re-
sponse maximum relative estimation error can be evaluated

Ĝ(jω) − G(jω)
E := max | |. (2.60)
ω∈(0, ωM ) G(jω)

And the following criterion will be used to validate the solution

E≤ (2.61)

where  is a user-specified fitting error threshold.  is specified according

to the desired accuracy of the RSL approximation to the original dynamics
G(s). Usually  may be set between 1% ∼ 10%. If (2.61) is met, the procedure
stops. Otherwise, one increases n by 1 until the smallest integer n such that
E ≤ .
Algorithm 2.4.1 Seek a reduced-order model Ĝ(s) of order n in (2.47) given
G(s) or G(jω), approximation threshold  and the initial time delay L0 and
parameter vector θ0 .
step 1. Choose N between 15 ∼ 20, set ∆L = 1.5L N , and obtain Lk = 0.5L0 +
0

k∆L, k = 0, 1, · · · , N .
step 2. Start from Ĝ(s) with n = 1.
step 3. For each Lk , find the nth order rational approximation solution ĝ0 (s)
to the modified process g0 (s) = G(s)eLk s with the RLS method in
(2.55) - (2.56) and evaluate the corresponding approximation error
E in (2.61) for Ĝ(s) = ĝ0 (s)e−Lk s .
 2.4 Model Reduction 49

step 4. Take Ĝ(s) as the solution if it yields the minimum error E and E ≤ .
Otherwise, n + 1 → n and go to Step 3.
The preservation of stability is a crucial issue in frequency domain based
model reduction. Let Ĝ(s) = ĝ0 (s)e−Ls be the one that yields the smallest
approximation error e in (2.48). It is noted that Algorithm 2.4.1 with zero ini-
tial parameter vector might result in unstable Ĝ(s), especially for high-order
models, even though G(s) is stable. One can use the stability tests and pro-
jection algorithm (Ljung and Soderstrom, 1983) to constrain the poles of the
model to be in the stable region. However, simulation shows that this method
can slow down the convergence of the recursive least square considerably and
may result in large modelling error, if the dynamics of the transfer function
to be modelled is complicated. Here, we notice that since (2.48) is a nonlin-
ear problem, Algorithm 2.4.1 may yield different local optimal solutions, if it
starts with different initial settings. Among those solutions, only stable Ĝ(s)
models are required given a stable G(s). If we initiate the algorithm with
a stable model, the algorithm is likely to reach a stable approximate upon
convergence.
When n = 1, we set the initial model for Ĝ(s) as

β0
Ĝ0 = e−L0 s . (2.62)
s + α0

Matching Ĝ0 (jω) to G(jω) at ω = 0 and ω = ωc , where ωc is the phase cross

over frequency of G(s), i.e., ∠G(jωc ) = −π, we get
 q

 α = ω |G(jωc )|2
 0 c G2 (0)−|G(jωc )|2 ,
β0 /α0 = G(0), (2.63)


L = 1 {− arg[G(jω )] − tan−1 ( ωc )},
0 ωc c α0

When n = 2, one may adopt the following structure,

β0
Ĝ0 = e−L0 s . (2.64)
s2 + α1 s + α0

Similarly, match Ĝ0 (jω) to G(jω) at the two points ω = ωb and ω = ωc , where
∠G(jωb ) = − π2 and ∠G(jωc ) = −π. It then follows (Wang et al., 1999a) that
the parameters β0 , α1 , α0 and L0 can be determined as
 sin(ωc L0 ) ωc |G(jωc )|

 cos(ωb L0 ) = ωb |G(jωb )| ,

 h i−1

 β0 = (ωc2 − ω 2 ) sin(ωb L0 ) + cos(ωc L0 )
b |G(jωb )| |G(jωc )| ,
sin(ωc L0 ) (2.65)

 α1 /β0 = ωc |G(jωc )| ,

 h i−1

 α /β = (ω 2 − ω 2 ) ωc2 sin(ωb L0 ) + ωb2 cos(ωc L0 )
0 0 c b |G(jωb )| |G(jωc )| .
50 2. Representations of Linear Dynamic Systems

For the cases of n > 2, we can set the initial model Ĝ0 (s)of the respective
orders with β0 , α1 , α0 and L0 determined as in (2.65), while all the remaining
high-degree coefficients are set to 0. Our extensive simulation shows that this
technique works very well.

Example 2.4.1. Consider a high-order plant (Maffezzoni and Rocco, 1997):

(−2.7s + 1)(158.5s2 + 6s + 1) −14s

G(s) = 2.15 e .
(17.5s + 1)4 (20s + 1)
With zero initial parameter vector, Algorithm 2.4.1 gives a reduced order
model
−0.0275s2 − 0.0010s − 0.0001 −63.20s
Ĝ(s) = e , (2.66)
s3 − 0.0129s2 − 0.0029s − 0.0001
which is unstable. When the parameters obtained from (2.65) are adopted
for the initial model, we get

0.4687s2 + 0.0038s + 0.0024

Ĝ(s) = e−41.90s , (2.67)
s3 + 1.2160s2 + 0.0665s + 0.0011
which is stable with E = 2.71%. The frequency responses of the actual and
estimated models are shown in Figure 2.1. ♦

0.5

−0.5

−1

−1.5

−2
−1 −0.5 0 0.5 1 1.5 2 2.5

Fig. 2.1. Frequency Domain Model Reduction

(—— actual process; ∗ ∗ ∗ unstable model; − ◦ − ◦ − stable model)
 2.4 Model Reduction 51

2.4.2 Stable Model Reduction

The method described in the preceding subsection cannot guarantee the

preservation of stability of the reduced-order models. This is a typical problem
associated with frequency domain methods. On the other hand, time domain
model reduction is easier to deal with stability preservation. One sees that
the step response of a stable G(s) will remain finite while that of an unstable
Ĝ(s) will tend to infinity. The squared error between them will diverge, and
unstable Ĝ(s) will be excluded from the solutions to the problem of minimiz-
ing such an error, in other words, the solution must be stable. Based on this
idea, we present a time domain model reduction algorithm. This algorithm
not only preserves stability but also enables non-iterative estimation of all
model parameters including time delay.

Construction of Step Response from G(s) or G(jω). Suppose that

plant is stable with G(s) or G(jω) given. Let the plant input u(t) be of step
type with size h. Then the plant output step response in Laplace domain is
h
Y (s) = G(s) . (2.68)
s
It seems very easy to obtain the corresponding time response y(t) by simply
applying the inverse Fourier transform (F −1 ). However, since the steady state
part of y(t) is not absolutely integrable, such a calculation is inapplicable and
meaningless (Wang et al., 1997b). To solve this obstacle, y(t) is decomposed
into

y(t) = y(∞) + ∆y(t) = G(0)h + ∆y(t).

Applying the Laplace transform to both sides gives

G(0)
Y (s) = + L{∆y(t)}. (2.69)
s
Bring (2.68) in, we have

G(s) − G(0)
L{∆y(t)} = h .
s
Applying the inverse Laplace transform yields

G(s) − G(0)
∆y(t) = hF −1 { }.
s
Thus, the plant step response is constructed as

G(jω) − G(0)
y(t) = h[G(0) + F −1 { }], (2.70)
jω
52 2. Representations of Linear Dynamic Systems

where F −1 may easily be implemented by the inverse fast Fourier transform

(IFFT). Alternatively, the step response y(t) may be obtained from a real
time step test on the given stable plant.

Identification from Step Response. Suppose that the given stable step
response is to be fitted into the following model:

b1 sn−1 + b2 sn−2 + · · · + bn−1 s + bn −Ls

Y (s) = Ĝ(s)U (s) = e U (s),
sn + a1 sn−1 + · · · + an−1 s + an
(2.71)

or equivalently by
y (n) (t) + a1 y (n−1) (t) + . . . + an−1 y (1) (t) + an y(t) = b1 u(n−1) (t − L)
+b2 u(n−2) (t − L) + . . . + bn−1 u(1) (t − L) + bn u(t − L), (2.72)
where L > 0. For an integer m > 1 , define
Z (m) Z tZ τm Z τ2
f= ··· f (τ1 )dτ1 · · · dτm .
[0,t] 0 0 0

Under zero initial conditions, for u(t) = h1(t), it follows that

Z (m)
1
u(t − L) = (t − L)m h.
[0,t] m!

Integrating (2.72) n times gives

Z (1) Z (2) Z (n−1) Z (n)
y(t) = −a1 y − a2 y · · · − an−1 y − an y
[0,t] [0,t] [0,t] [0,t]
1 1
+hb1 (t − L) + hb2 (t − L)2 + · · · + hbn−1 (t − L)n−1
2 (n − 1)!
1
+ hbn (t − L)n ,
n!
Z (1) Z (n) Xn
bj (−L)j
= −a1 y − · · · − an y + ht0
[0,t] [0,t] j=1
j!
n
X n
bj (−L)j−1 ht2 X bj (−L)j−2
+ht1 + + ···
j=1
(j − 1)! 2! j=2 (j − 2)!
n
htn−1 X bj (−L)j−(n−1) htn
+ + bn . (2.73)
(n − 1)! j=n−1 1! n!

Define
 2.4 Model Reduction 53



 γ(t) = y(t),
h R i

 (1) R (n)

 φT (t) = − [0,t] y, · · · , − [0,t] y, h, ht, ht2 tn h
h 2! , · · · , n! ,
T
Pn b (−L)j Pn bj (−L)j−1 (2.74)

 θ = a1 , · · · , an , j=1 j j! , j=1 (j−1)! , · · · ,

 Pn i

 j−(n−1)
j=n−1 bj (−L) , bn .

Then (2.73) can be expressed as

γ(t) = φT (t)θ, if t > L.
Choose t = ti and L 6 t1 < t2 < · · · < tN . Then
Γ = Φθ, (2.75)

where Γ = [γ(t1 ), γ(t2 ), ..., γ(tN )]T , and Φ = [φ(t1 ), φ(t2 ), ...., φ(tN )]T .
In (2.75), the parameter vector θ which minimizes the following error,
min(Γ − Φθ)T (Γ − Φθ),
θ

is given by the least square solution:

θ̂ = (ΦT Φ)−1 ΦT Γ. (2.76)
In practice, the true plant output ŷ may be corrupted by computational
errors or measurement noise v(t), and
y = ŷ + v, (2.77)
where v is supposed to be a strictly stationary stochastic process with zero
mean. In this case, (2.75) is again modified to be
Γ = Φθ + ∆, (2.78)

where ∆ = [δ1 , δ2 , · · · , δN ]T , and

Z (1) Z (2) Z (n)
δi = [−v(ti ) − a1 v − a2 v ··· − an v].
[0,ti ] [0,ti ] [0,ti ]

To get an unbiased estimate, the instrumental variable matrix Z is con-

structed as
 1 1 1 
tn+1 tn+2
· · · t2n 1 t1 · · · tn1
1 1 1
 n+11 1 1
· · · t2n 1 t2 · · · tn2 
 t2 tn+2 
Z= 
2
.. . . .. .. .. . . ..  , (2.79)
2
 ..
 . . . . . . . . 
1 1 1
tn+1 tn+2
· · · t2n 1 tN · · · tnN
N N N

and the estimate is then given by

θ̂ = (Z T Φ)−1 Z T Γ. (2.80)
54 2. Representations of Linear Dynamic Systems

Theorem 2.4.1. For a linear time-invariant process of n-th or higher-order,

if the noise in the output measurement is a zero-mean strictly stationary
stochastic process and the number of output samples N satisfies N > 2n + 1,
then the estimate given by

θ̂ = (Z T Φ)−1 Z T Γ (2.81)

is consistent, where Z is given in (2.79). If there is no noise, Z may be replaced

by Φ.

Once θ is estimated from (2.80), one has to recover the process model
coefficients: L, ai and bi , i = 1, 2, · · · , n. It follows from (2.74) that

ak = θk k = 1, 2, · · · , n;

and
 
(−L)2 (−L)3 (−L)n−1 (−L)n  
−L ···
 2! 3! (n−1)! n!  θn+1
 1 −L (−L)
2
··· (−L)n−2 (−L)n−1  b1
 


2! (n−2)! (n−1)!     θn+2 
b2
 (−L)n−3 (−L)n−2     θn+3 
 0 1 −L · · · (n−3)! (n−2)!    
b3 
 n−3     θn+4 
 0 0 1 ··· (−L) n−4
(−L)
 =.. . (2.82)
 (n−4)! (n−3)!     .. 
.
 .. .. .. .. .. ..    . 
 . . . . . .  bn−1  
 θ2n 

 
 0 0 0 ··· 1 −L  bn θ2n+1
0 0 0 ··· 0 1

The last n rows of (2.82) give bi in terms of L:

   (L)n−2 (L)n−1
 θ 
L2 n+2
b1 1 L · · ·
2! (n−2)! (n−1)!  θn+3 
 b2    0 1 L · · · (L)n−3 (L)n−2 
  
   (n−3)! (n−2)!   θn+4 
 ..     
 .  =  .. .. .. . . ... ...  ·  ..  . (2.83)
  . . . . 
  . 
bn−1   
0 0 0 ··· 1 L   θ2n 
bn 0 0 0 ··· 0 1 θ2n+1

Substituting (2.83) into the first row of (2.82) yields the following n-degree
polynomial equation in L:
n
X θn+i+1
Li = 0. (2.84)
i=0
i!

Equation (2.84) has n roots for L. In selecting a suitable solution, a rule of

thumb is to choose one that leads to the minimal output error between the
step response of the estimated model and the given one. Once L is determined,
bi can be easily computed from (2.83).
 2.4 Model Reduction 55

Algorithm 2.4.2 Seek a reduced-order model Ĝ(s) of order n in (2.71) given

G(s) or G(jω).
step 1. Construct the output step response y(t) from G(s) or G(jω) with the
IFFT by (2.70);
step 2. Compute the regression vector θ̂ by (2.76) or (2.80);
step 3. Recover the model parameters from θ̂ by (2.76) and (2.83) and (2.84).
Some special cases can simplify the formulas. For first-order plus dead
time (FOPDT) modelling:
b −Ls
G(s) = e ,
s+a
(2.74) becomes


 γ(t) = y(t),
h R i
 t
φT (t) = − 0 y(τ )dτ, h, ht ,

 h i
θT = a, −bL, b ,

and the model parameters are recovered from

 T h iT
a, b, L = θ1 , θ3 , − θθ31 .

For second-order plus dead time (SOPDT) modelling:

b1 s + b2
G(s) = e−Ls ,
s2 + a1 s + a2
(2.74) becomes


 γ(t) = y(t),
h R i
 t RtRτ
φT (t) = − 0 y(τ )dτ, − 0 0 y(τ1 )dτ1 dτ, h, th, 12 t2 h,

 h i
θT = a , a , −b L + 1 b L2 , b − b L, b ,
1 2 1 2 2 1 2 2

and the model parameters are recovered from

 T h iT
a1 , a2 , b1 , b2 , L = θ1 , θ2 , β, θ5 , − −θθ45+β

with
( p
− θ2 − 2θ5 θ3 , if a inverse response is detected;
β= p 24
θ4 − 2θ5 θ3 , otherwise.

A few parameters need to be specified by users before Algorithm 2.4.2 is

applied. The guidelines for their selection are given now, which can lead to
effective implementation of this algorithm.
56 2. Representations of Linear Dynamic Systems

Choice of t1 . It is noted from the above development that the first

sample y(t1 ) should not be taken into the algorithm until t1 > L, when the
output deviates from the previous steady state. In practice, the selection of
the logged y(t) after ti > L can be made as follows. Before the step input
is applied, the output will be monitored for a period called the ‘listening
period’, during which the noise band Bn can be found. After a step change
in the input is applied, the time t at which y(t) satisfies
abs(mean(y(t − τ, t))) > 2Bn
is considered to meet t > L, where τ is user-specified time interval and is
used for averaging.
Choice of tN . The initial part of the step response contains more high
frequency information, while the part of the response after the steady state
contains only zero frequency information, i.e., at ω = 0. Extensive simulation
suggests that tN be set at 1.2 ∼ 1.5Tset , where Tset is the settling time,
defined as the time required for the output to settle within ±2% of its steady
state.
Choice of N . The computational effort becomes heavy if too many
samples are taken into consideration, which leads to the large size of Φ.
Moreover, ΦT Φ or Z T Φ may tend to become ill-conditioned for very large N
and this may cause the computational difficulty in estimating θ̂. Therefore,
N has to be limited. For the case with a large number of recorded data, the
default value of N is recommended to be 200, and ti may be set as
i−1
ti = t1 + (tN − t1 ), i = 1, 2, · · · , N.
N
For a better assessment of model accuracy, identification errors in both
time domain and frequency domain are used. The time domain identification
error is measured over the transient period by standard deviation:
N
1 X
ε= [y(ti ) − ŷ(ti )]2 , (2.85)
N i=1

where y(ti ) and ŷ(ti ) are the step responses of the plant G(s) and its model
Ĝ(s), respectively, under the same step input. Once the identification is car-
ried out, the model Ĝ(s) is available and the frequency error E in (2.60) is
measured as the worst-case error. Here, the frequency range [0, ωc ] is con-
sidered, where ∠G(jωc ) = −π, since this range is the most significant for
control design. To show the robustness of the proposed method, noise is in-
troduced into the output response. In the context of system identification,
noise-to-signal ratio defined by
mean(abs(noise))
N SR = .
mean(abs(signal))
is used to represent noise level.
 2.5 Conversions between Continuous and Discrete Systems 57

Example 2.4.2. Reconsider the high-order plant in Example 2.4.1. The out-
put step response is first constructed using the IFFT. The area method
(Rake, 1980) gives the model:
2.15
Ĝ(s) = e−53.90s ,
46.69s + 1
with  = 1.15 × 10−2 and E = 60.87%. The FOPDT model estimated by
Algorithm 2.4.2 is
0.0396 −49.9839s
Ĝ(s) = e ,
s + 0.0184
with  = 8.6 × 10−3 and E = 48.12%. Our SOPDT model is
0.0011
Ĝ(s) = e−28.8861s ,
s2 + 0.0343s + 0.0005
with  = 4.0631 × 10−4 and E = 5.81%. For third-order plus dead time model
(TOPDT), we have

0.2059s2 + 0.0005s + 0.0001

Ĝ(s) = e−31.9265s ,
s3 + 0.9969s2 + 0.0382s + 0.0001
with  = 2.0009 × 10−4 and E = 1.27%. It can be seen that the results
from Algorithm 2.4.2 are consistently better than that of the area method.
The estimation errors decrease with model order n. Especially from FOPDT
to SOPDT, errors decrease dramatically, but the error decrease slows down
from SOPDT to TOPDT. In this case, an SOPDT model is good enough.
Step and frequency responses of the actual and estimated models are shown
in Figure 2.2. ♦

In this section, we have discussed two model reduction approaches. Re-

cursive Lease Square (RLS) is simple to employ, and stability can be taken
care of but cannot be guaranteed. The time domain method involves several
steps but can preserve stability.

2.5 Conversions between Continuous and Discrete

Systems

This book will address control system design for continuous-time systems
only. For computer implementation, the designed continuous controllers have
to be converted to discrete forms. In general, signals from computer related
applications are often manipulated in digital form. Conversion between con-
tinuous and discrete time systems is thus needed in these applications. The
conversion accuracy is important and directly affects the final performance of
58 2. Representations of Linear Dynamic Systems

2.5

1.5

0.5

−0.5
0 50 100 150 200 250 300 350

(a) T ime responses

0.5

−0.5

−1

−1.5

−2
−1.5 −1 −0.5 0 0.5 1 1.5 2 2.5

(b) F requency responses

Fig. 2.2. Time Domain Model Reduction

(—— actual process; · · · area method; − · − · −· proposed FOPDT;− − − proposed
SOPDT; * * * proposed TOPDT)
 2.5 Conversions between Continuous and Discrete Systems 59

the system. In this section, popular discrete equivalents to continuous transfer

functions are first reviewed. Then a frequency response fitting method based
on Recursive Least Square (RLS) to solve the model conversion problem is
presented. The objective of this method is to obtain the discrete transfer
function (DTF) so that its frequency response fits that of the continuous
transfer function (CTF) as well as possible. A DTF model can be obtained
either from a CTF or directly from the system frequency response such as
Bode plot data.

2.5.1 Popular Rules

We consider this problem: Given a continuous time transfer function G(s),

what discrete time transfer function G(z) will approximately have the same
characteristics?

Discrete equivalents by numerical integration. We recast a dynamic

system G(s) to the corresponding differential equation, numerically integrate
with the step size T the equation to get the discrete solution. Different numer-
ical integration rules will yield different solutions. Most popular are forward,
backward and Trapezoid (Bilinear) rules. Applying z-transform to discretized
systems will get respective G(z). The results are summarized in Table. 2.1.
Note that the conversions can be in two ways, i.e, either from G(s) to G(z),
or from G(z) to G(s). To know stability preservation properties of these rules,
we let s = jω in these equations and obtain the boundaries of the regions
in the z-plane. The shaded areas in Figure 2.3 corresponds to the stability
region in s-plane (Re s < 0).

Table 2.1. Integration Based Rules

Method G(s) → G(z) G(z) → G(s)
Forward Rule s = z−1
T
z = 1 + Ts
Backward Rule s = z−1
Tz
1
z = 1−T s
1+T s/2
Bilinear Rule s = T2z−1
z+1
z = 1−T s/2

Zero-pole Mapping Equivalents. A very simple but effective method

of obtaining a discrete equivalent to a continuous G(s) is to relate poles of
G(s) to the poles of G(z) according to z = esT . The same mapping is applied
to zeros, too. The following heuristic rules are used to locate poles and zeros
of G(z).
(i) if G(s) has a pole at s = −a, then G(z) has a pole at z = e−aT . If G(s)
has a pole at −a + jb, then G(z) has a pole at rejθ , where r = e−aT and
θ = bT .
60 2. Representations of Linear Dynamic Systems

Unite circle

(a) (b) (c)

Fig. 2.3. Maps of the left-half s-plane to the z-plane by the integration rules in
Table .2.1. Stable s-plane poles map into the shaded regions in the z-plane. The
unite circle is shown for reference. (a) Forward rectangular rule. (b) Backward
rectangular rule.(c) Trapezoid or bilinear rule.

(ii) All finite zeros are mapped in the same way.

(iii) The zeros of G(s) at s = ∞ are mapped in G(z) to the point z = −1. or
(iiia) One zero of G(s) at s = ∞ is mapped into z = ∞.
(iv) There holds gain matching, i.e.

G(s)|s=0 = G(z)|z=1 .

Application of these rules to G(s) = a/(s + a) gives

(z + 1)(1 − e−aT )
G(z) = .
2(z − e−aT )

or using (iiia), we get

1 − e−aT
G(z) = .
z − e−aT
Hold Equivalents. In this case, we want to design a discrete system G(z)
that, with an input consisting of samples of a continuous signal u(t), has an
output which approximates the output of the continuous system G(z) whose
input is the continuous signal u(t). We can do so through hold operation. If
the hold is of the zero-order, the conversion is given by

G(s)
G(z) = (1 − z −1 )Z{ }.
s
Suppose again
a
G(s) = .
s+a
 2.5 Conversions between Continuous and Discrete Systems 61

It follows that
G(s) a 1 1
Z{ } = Z{ } = Z{ } − Z{ }
s s(s + a) s s+a
(1 − e−aT z −1 ) − (1 − z −1 )
= ,
(1 − z −1 )(1 − e−aT z −1 )
so that
1 − e−aT
G(z) = .
z − e−aT

2.5.2 A Frequency Response Fitting Approach

The rules discussed above are applicable mainly to rational G(s) only. We
now develop a conversion technique which is suitable for non-rational G(s),
too. Let a continuous transfer function Gc (s) or frequency response Gc (jω)
be given. An equivalent discrete transfer function

βm z m + βm−1 z m−1 + ... + β1 z + β0

Ĝd (z) = , (2.86)
z m + αm−1 z m−1 + ... + α1 z + α0

is to be found to match the frequency response of Ĝd (z) to Gc (s) with respect
to the map,

z = eT s . (2.87)

This problem is formulated as

M
X
minJ , min |Gc (jωi ) − Ĝd (ejT ωi )|2 , (2.88)
Ĝd Ĝd
i=1

where (ω1 , ωM ) defines the frequency range of interest. Equation (2.88) falls
into the framework of transfer function identification in frequency domain, in
which Gc is the known frequency response while Ĝd is a parametric transfer
function to be identified. Obviously, the RLS based method (Algorithm 2.4.1)
in the preceding section is immediately applicable. To make the best use of
it, two minor remarks are given as follows.
The frequency range in the optimal fitting is a useful tuning parameter.
In general, our studies suggest that the frequency range (ω1 · · · ωM ) be cho-
1 1 1
sen as ( 100 ωb , ωb ) with the step of ( 100 ∼ 10 )ωb , where ωb is the bandwidth
of Gc (jω). In this range, RLS yields satisfactory fitting results in frequency
domain. Next, the preservation of stability is again a crucial issue. We adopt
the same strategy, i.e., make the initial model stable. But, the formulas in
(2.62) and (2.64) are no longer useful as they only ensure stability of contin-
uous systems. To get a stable initial Gd0 (z) from the given stable Gc (s), note
from Figure 2.3(a) that the Bilinear rule preserves stability under continuous
62 2. Representations of Linear Dynamic Systems

and discrete conversions. Thus, for a rational Gc (s), we set the initial model
Gd0 (z) according to the Bilinear rule. For other Gc (s), the idea behind (2.62)
and (2.64) can still be used, that is, match Gd0 (z) to Gc (s) at two selected
frequencies to get a stable Gd0 (z). Our extensive simulation shows that this
technique works very well.
The method described above is also applicable to inverse conversion, i.e.,
from DTF to CTF, with obvious changes of the rules of continuous and
discrete frequency responses in (2.88).
We now illustrate the above technique with some typical plants. In sim-
ulation, the sampling time T is chosen to meet the standard rule:
ωs
6. . 40, (2.89)
ωb
where ωs = 2πT is the sampling frequency. Once a discrete equivalent is found,
the frequency response maximum relative error similar to (2.60) is used to
assess the conversion accuracy.
Example 2.5.1. Consider the continuous filter transfer function (Franklin et
al., 1990, pp. 141):
1
Gc (s) = ,
s3 + 2s2+ 2s + 1
ωs
with unity pass bandwidth. For T = 1 sec, i.e., ωb = 2π > 6, the Bilinear
rule gives
0.0476z 3 + 0.1429z 2 + 0.1429z + 0.0476
Ĝd (z) = ,
z 3 − 1.1900z 2 + 0.7143z − 0.1429
with the error E = 24.67 %, while the zero-pole mapping transformation does
0.0920(z + 1)2
Ĝd (z) = ,
(z − 0.3679)(z 2 − 0.7859z + 0.3679)
with E = 62.84 %. For m = 3, the fitting method yields
0.0080z 3 + 0.2257z 2 + 0.1381z − 0.0072
Ĝd (z) = ,
z 3 − 1.1583z 2 + 0.6597z − 0.1367
with E = 0.26 %. In Figure 2.4, the responses of three models are compared,
and the fitting method yields much better performance than others. More-
over, Figure 2.5 shows that in the reasonable range of ωs given in (2.89),
the fitting method exhibits a consistent superior approximation performance
than Bilinear and zero-pole mapping schemes. ♦
Example 2.5.2. Consider the system with time delay
s−1
Gc (s) = e−0.35s , (2.90)
s2 + 4s + 5
 2.5 Conversions between Continuous and Discrete Systems 63

0.8

0.6
gain

0.4

0.2

0 0.5 1 1.5 2 2.5

frequency w (rad/s)

−50

−100
phase (°)

−150

−200

−250

−300

0 0.5 1 1.5 2 2.5

frequency w (rad/s)

Fig. 2.4. Digital equivalents to 3rd-order lowpass filter

(—— CTF, · · · ∗ · · · Fitting, · · · o · · · Bilinear, · · · + · · · P-Z mapping)

with T = 0.1 sec (Control System Toolbox User’s Guide, MathWorks, ver. 4.2,
1998, c2d function). The Toolbox, via Triangle approximation, gives

0.0115z 3 + 0.0456z 2 − 0.0562z − 0.0091

Ĝd (z) = z −3 ,
z 3 − 1.6290z 2 + 0.6703z
with E = 5.12%. The fitting method generates

0.0042z 2 + 0.0684z − 0.0807

Ĝd (z) = z −3 ,
z 2 − 1.6333z + 0.6740
with E = 0.50%, a much improved accuracy. ♦

To end this section, it is worth mentioning that the fitting method is suit-
able for general systems with either parametric or non-parametric source
models, and can achieve high performance with good chance of stability
preservation. In fact, it can be also applied to many other cases such as
determining optimal sampling time, optimal model order, and re-sampling a
discrete systems. Readers are referred to Wang and Yang (2001) for details.
64 2. Representations of Linear Dynamic Systems

0
10
E

−1
10

−2
10

−3
10

−4
10

−5
10

−6
10
10 15 20 25 30 35 40
ws/wb

Fig. 2.5. Approximation error E versus sample rate ωs

(—— Fitting, − · − · − Bilinear, · · · · · · P-Z mapping)

2.6 Notes and References

The standard linear system theory is well covered in Kailath (1980) and
Chen (1984). The Polynomial matrix approach to linear multivariable system
analysis and design became popular since 1970’s (Rosenbrock, 1974; Callier
and Desoer, 1982; Wolovich, 1978; Kucera, 1979). For model reduction, the
most popular are the balanced truncation method (Moore, 1981; Pernebo
and Silverman, 1982; Enns, 1984), and the Hankel norm approximation
method (Glover, 1984). Optimization-based approaches to model reduction
are discussed in Meier and Luenberger (1967), Hyland and Bernstein (1985),
Baratchart et al. (1991), Spanos et al. (1992), Zhou (1995). But the prob-
lem for delay systems is less developed and stability of reduced-order models
usually cannot be guaranteed. The algorithms given in Section 4 are from
Yang (2001). Standard formulas for converting a continuous time model to
a discrete-time one are well known (Kuo, 1980; Rattan, 1984; Franklin et
al., 1990). And optimization based discretization has been widely studied
by Chen and Francis (1991), Chen and Francis (1995), Keller and Anderson
(1992), Shieh et al. (1992) and Rafee et al. (1997), which are only applicable
for delay-free cases. However, the presence of time delay makes a lot of dif-
ference and stability is a crucial issue. The method in Section 5 is based on
Wang and Yang (2001).
3. Stability and Robustness of Feedback
Systems

Feedback is the main tool for control systems. Stability and robustness are
two key issues associated with any feedback design, and are the topics of this
chapter. Section 1 addresses internal stability of general interconnected sys-
tems and derives a powerful stability condition which is applicable to systems
with any feedback and/or feedforward combinations. Section 2 focuses on the
conventional unity output feedback configuration and gives the simplifying
conditions and Nyquest-like criteria. The plant uncertainties and feedback
system robustness to them are discussed in Section 3, both structured and
unstructured perturbations are introduced. A special attention is paid to the
case where the phase perturbation is limited, a realistic situation.

3.1 Internal Stability

Stability is a fundamental requirement for any feedback systems. In the state-

space setting, all signals inside the system can be expressed as a linear com-
bination of its state variables and they are bounded for any bounded inputs
if and only if all state variables of the system are bounded for any bounded
inputs. However, state-space descriptions of systems may not be available in
many cases. In this section, internal stability of interconnected systems is
considered from an input-output setting. It is shown that a system consisting
only of single-input
Q and single-output (SISO) plants is internally stable if and
only if ∆ i pi (s) has all its roots in the open left half of the complex plane,
where pi (s) are the denominators of the plant transfer functions and ∆ is the
system determinant as defined in the Mason’s formula. A general theorem
is also presented which is applicable to the case where the system may have
multi-input and/or multi-output plants. Several typical control schemes are
employed as illustrative examples to demonstrate the simplicity and useful-
ness of these results in internal stability analysis and stabilization synthesis.

3.1.1 Stability Criterions

Let a general interconnected system have n plants with proper transfer

function matrices Gi (s), i = 1, 2, · · · , n. Assume that for each i, Gi (s) =

Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 65-114, 2003.

Springer-Verlag Berlin Heidelberg 2003
66 3. Stability and Robustness of Feedback Systems

A−1
i (s)Bi (s)ÿ is a
left coprime polynomial matrix fraction (PMF). Then
pi (s) = det Ai (s) is the characteristic polynomial of Gi (s). Gi (s) can be
considered (Rosenbrock 1974) to arise from the equations:
Ai (s)ξi (s) = Bi (s)ui (s), (3.1a)
yi (s) = ξi (s), (3.1b)
where ui and yi are the input and output vectors of plant Gi (s) respectively.
n plants in form of (3.1) may be combined to
A(s)ξ(s) = B(s)U (s), (3.2a)
Y (s) = ξ(s), (3.2b)
where
A(s) = block diag{A1 (s), A2 (s), · · · , An (s)},
B(s) = block diag{B1 (s), B2 (s), · · · , Bn (s)},
 T
ξ = ξ1T ξ2T · · · ξnT ,
 T
U = uT1 uT2 · · · uTn .
In an interconnected system, the outputs of some plants may be injected into
other plants as inputs. Besides, various exogenous signals (references and
disturbances) rj , j = 1, 2, · · · , m, may also come into the system. Thus, U
has the general form:

U (s) = CY (s) + DR(s), (3.3)

T
where R = [r1 r2 · · · rm ] , and C represents the system internal input-
output connection between plants, and C and D can be easily read out by
inspection of the plants’ interconnections.
Combining (3.2) and (3.3) yields

(A(s) − B(s)C)ξ(s) = B(s)DR(s). (3.4)

Definition 3.1.1. The interconnected system described in (3.2) and (3.3)

is called internally stable if and only if the system characteristic polynomial
pc (s) defined by

pc (s) = det(A(s) − B(s)C) (3.5)

has all its roots in the open left half of the complex plane.

Let G(s) = A−1 (s)B(s) = block diag{G1 , G2 , · · · , Gn }. To proceed with

stability analysis, we need a well-posedness assumption.

Assumption 3.1.1 det(I − G(∞)C) 6= 0.

 3.1 Internal Stability 67

This assumption ensures the properness of all elements of the transfer

function matrix from the exogenous signal vector R(s) to the internal signal
vector ξ(s). To see this, one notes that A(s) − B(s)C = A(s)(I − G(s)C) is
nonsingular under Assumption 3.1.1, and has from (3.4)

ξ(s) = (A(s) − B(s)C)−1 B(s)DR(s)

= (I − G(s)C)−1 (s)G(s)DR(s).

The transfer matrix between ξ(s) and R(s) meets

 
lim (I − G(s)C)−1 G(s)D = (I − G(∞)C)−1 G(∞)D < ∞
s→∞

for proper G(s). Obviously, Assumption 3.1.1 is true if G(s) is strictly proper,
i.e., all plants Gi (s) are strictly proper.

Fig. 3.1. System well-posedness

Assumption 3.1.1 is necessary to rule out any possibility of improper

closed-loop transfer functions resulting from proper G. For example, the sys-
tem in Figure 3.1 yields an improper transfer function,
1−s
s
HY R = = 1 − s,
1 + 1−s
s

though G(s) = 1−s s is proper. For this case, one sees that U = R − Y and
C = −1, which leads to 1 − G(s)C = 1 + G = 1s and det(1 − G(∞)C) = 0,
violating Assumption 3.1.1.
Recall that a rational function is stable if it is proper and all its poles have
negative real parts. The properness has to be included, otherwise, a bounded
input may generate an unbounded output even though all the poles of the
rational function have negative real parts. Take the above example again for
instance, one has pc (s) = det(A(s) − B(s)C) = (s + 1 − s) = 1 for which there
is no pole and thus no unstable pole. But HY R = 1 − s, for the bounded
input, say, r(t) = sin(t2 ), the output response under zero initial condition is
y(t) = sin(t2 ) − 2t cos(t2 ) and unbounded.
68 3. Stability and Robustness of Feedback Systems

Refer back to the interconnected system described in (3.2) and (3.3), for
ease of future reference, we define the system connection matrix as

W (s) = I − G(s)C, (3.6)

and pc (s) in (3.5) can be rewritten as

pc (s) = det(A(s))det(I − A−1 (s)B(s)C)
Yn
= det(I − G(s)C) pi (s)
i=1
n
Y
= det(W (s)) pi (s). (3.7)
i=1

Hence, we have the following result.

Theorem 3.1.1. Under Assumption 3.1.1, a linear time-invariant inter-

connected system is internally stable if and only if pc (s) in (3.7) has all its
roots in the open left half of the complex plane.

The system which Theorem 3.1.1 can be applied to can contain multi-
input plants. In such a case, it is here called the system with vector signals
since its input and/or output are vector signals. On the other hand, if a lin-
ear time-invariant interconnected system consists only of SISO plants Gi (s),
then such a system is here called the system with scalar signals because all
the signals in the system is scalar. For such systems, we are going to de-
velop an elegant stability criterion which makes use of plants’ characteristic
polynomials and the system determinant only without evaluating det W (s)
(which may be difficult to calculate for large-scale systems). This criterion
may also be applied to systems with vector signals if all multi-input and/or
multi-output plants involved in the system are viewed to consist of several
SISO subsystems and are represented by their coprime polynomial matrix
fractions (PMFs) or minimal state-space realizations in the signal flow graph
of the system.
For a system with scalar signals, we first draw its signal flow graph. Recall
that a signal flow graph consists of nodes and directed branches. A node per-
forms the functions of addition of all incoming signals and then transmission
of it to all outgoing branches. A branch connected between two nodes acts
as a one-way signal multiplier and the multiplication factor is the transfer
function of the corresponding plant. A loop is a closed path that starts at a
node and ends at the same node. We say that a signal goes through a loop
if and only if it passes through all nodes and branches in the loop once. If
two loops have neither common nodes nor branches, then they are said to be
nontouched. The loop gain for a loop is the product of all transfer functions
of the plants in the loop. The system determinant ∆ is defined (Mason 1956)
by
 3.1 Internal Stability 69

X X X
∆=1− F1i + F2j − F3k + · · · , (3.8)
i j k

where F1i are the loop gains, F2j are the products of 2 non-touching loop
gains, F3k are products of 3 non-touching loop gains, · · · . The following is
the major result of this section.
Theorem 3.1.2. Let an interconnected system consist of n SISO plants
Gi (s), i = 1, 2, · · · , n, and the characteristic polynomial of Gi (s) be pi (s)
(simply the transfer function denominator) for each i. Suppose ∆(∞) 6= 0.
Then, the system is internally stable if and only if
n
Y
pc (s) = ∆(s) pi (s) (3.9)
i=1

has all its roots in the open left half of the complex plane.
Some insights may be drawn from the above theorem. Firstly, any pole-
zero cancellations will be revealed by pc (s) in (3.9). To see this, let p(s) be a
least commonQdenominator of all the loop gains after pole-zero cancellations
and po (s) = i pi (s). Then p(s) divides po (s) and the quotient contains the
poles of some plants that have been cancelled by other plant zeros but still
appear in pc (s) as the closed-loop poles. Secondly, one may see that any
unstable poles in the plants outside of the loop will be retained as the roots
of pc (s).
It should be pointed out that Theorem 3.1.1 can actually be applied to
systems with scalar signals too, since the scalar signal is a special case of a
vector signal. However, Theorem 3.1.2 is recommended in the case of scalar
signals as it is much simpler to use than Theorem 3.1.1. The internal stability
of several typical control schemes is now investigated with the help of these
results.

r + u1 y1 u2 y2 un yn
G1 G2 Gn
_

Fig. 3.2. Single-loop Feedback System

Example 3.1.1. Consider the single loop feedback system shown in Figure
3.2. It follows that
70 3. Stability and Robustness of Feedback Systems

 
0 0 0 ··· −I
I 0 0 ··· 0 
 
 0 
W = I − GC = I − diag{Gi }  0 I 0 ··· 
 .. .. .. . . .. 
. . . . . 
0 0 0 ··· 0
 
I 0 0 · · · G1
 −G2 I 0 ··· 0 
 
 ··· 0 
=  0 −G3 I .
 .. .. .. . . .. 
 . . . . . 
0 0 0 ··· I
Post-multiplying the i-th column by Gi Gi−1 · · · G1 and subtracting it from
the n-th column, i = 1, 2, · · · , n − 1, yields
 
I 0 0 ··· 0
 −G2 I 0 · · · G2 G1 
 
 
det(W ) = det  0 −G3 I · · · 0 
 .. .. .. . . .. 
 . . . . . 
0 0 0 ··· I
= ···
= det[I + Gn Gn−1 · · · G1 ].
Thus, it follows from Theorem 3.1.1 that Qnthe system is internally stable if and
only if pc (s) = det(I + Gn Gn−1 · · · G1 ) i=1 pi (s) has all its roots in the open
left half of the complex plane, where pi (s) are the characteristic polynomials
of Gi (s) respectively. ♦

By using a different indexing on Gi or the matrix identity:

 
I Y
det = det(I + XY ) = det[I + Y X]
−X I

for compatible X and Y , we obtain

Qn
pc (s) = det[I + Gn Gn−1 · · · G1 ] Qi=1 pi (s)
n
= det[I + Gn−1 · · · G1 Gn ] i=1 pi (s)
··· Qn
= det[I + G1 Gn · · · G2 ] i=1 pi (s).

Example 3.1.2. Consider the Smith predictor control system depicted in Fig-
ure 3.3. Assume that all the plants in the system are of SISO. Let g(s) =
g0 (s)e−Ls , g0 (s) = b(s)/a(s) and k(s) = β(s)/α(s) and be coprime polyno-
mial fractions. The system consists of 4 plants, g(s), g0 (s), e−Ls and k(s).
Their characteristic polynomials are a(s), a(s), 1 and α(s) respectively, so
Q4
that i=1 pi (s) = a2 (s)α(s). To find the system determinant ∆(s), one easily
 3.1 Internal Stability 71

r + y
k (s ) g (s )
_
_ +
g 0 (s) e - Ls
+
+

Fig. 3.3. Smith Predictor Control System

sees that the system has three loops which pass through (k, g0 ), (k, g0 , e−Ls )
and (k, g) respectively. Note that all these loops have the common branch k(s)
and thus they are touched. It follows that ∆ = 1+kg0 +kg−kg0 e−Ls = 1+kg0 .
According to (3.9), pc (s) for this Smith system is
4
Y ÿ

pc (s) = ∆(s) pi (s) = b(s)β(s) + a(s)α(s) a(s).
i=1
ÿ

It is well known that b(s)β(s) + a(s)α(s) can be assigned to an arbitrary
stable polynomial by a proper choice of α(s) and β(s) in k(s) for coprime
a(s) and b(s). Hence, the system is internally stabilizable if and only if a(s)
is a stable polynomial, or if and only if the process g(s) is stable. ♦
Example 3.1.3. Consider the input-output feedback system shown in Figure
3.4. Let g(s) = b(s)/a(s) be a coprime polynomial fraction, pf (s) and q(s) be
two given polynomials, and f (s) = a(s) − pf (s). Polynomials k(s) and h(s)
are determined by solving the polynomial equation:
k(s)a(s) + h(s)b(s) = q(s)f (s).
To analyze the system with our method, note that there are two touched
loops in Figure 3.4 whose gains are k(s)/q(s) and g(s)h(s)/q(s). It follows
that
 
k(s) h(s)
pc (s) = 1 − − g(s) q(a)a(s)
q(s) q(s)
= q(s)a(s) − k(s)a(s) − h(s)b(s) = q(s)pf (s),
where both q(s) and pf (s) are specified by the user. Thus, all the poles of
the system can be assigned arbitrarily. ♦

3.1.2 Proof of Theorem 3.1.2

For a system with scalar signals, all plants in the system have SISO transfer
functions. To emphasize this SISO property, we replace Gi (s) with gi (s).
72 3. Stability and Robustness of Feedback Systems

v + u y
g(s)
+

+ +
k(s) h(s)

1
q(s)

Fig. 3.4. Pole-assignment Control System

Consider the matrix C which represents the input-output connection between

plants. If the output of some plant is directly fed back as one of its inputs,
we view this feedback path as an additional plant with the unity gain. With
this convention, no output of a plant is an input to itself and thus cii = 0 for
the matrix C . It follows that C now reads as

 0, i = j;
cij = 1, i 6= j, and gj0 s output is an input to gi ; (3.10)

0, i 6= j, and gj0 s output is not an input to gi .

As a result, the system interconnection matrix W (s) = I − diag{Gi (s)}C

is determined as

 1, i = j;
wij = −gi (s), i 6= j, and gj0 s output is an input to gi ; (3.11)

0, i 6= j, and gj0 s output is not an input to gi .

which, in fact, can be obtained directly by inspection of the system without

referring to C at all.
What remains for proof of Theorem 3.1.2 is to show that for W given in
(3.11) there holds det(W (s)) = α∆(s) for a constant α. Obviously, det(W )

g7 g4

g1 g2
g3

g5

g6

Fig. 3.5. Signal Flow Graph of a Feedback System

 3.1 Internal Stability 73

involves matrix calculations while ∆ deals with the system signal flow graph.
To link them together, we will perform a series of expansions on det(W ) and
express det(W ) in terms of those items which have clear meanings on the
signal flow graph. It should be pointed out that the method we will use to
prove Theorem 3.1.2 below is general and applicable to any interconnected
system. But for ease of understanding, we shall frequently refer to an example
for demonstration when we introduce notations, concepts, derivations and
results. The example is the signal flow graph in Figure 3.5, and according to
(3.11) its system connection matrix W easily reads as
 
1 0 0 0 0 −g1 −g1
 −g2 1 0 0 −g2 0 0 
 
 0 −g3 1 −g3 0 0 0 
 
W =  0 0 −g4 1 0 0 0 . (3.12)
 0 −g5 0 −g5 1 0 0 
 
 0 0 −g6 0 0 1 0 
−g7 0 0 0 −g7 0 1

Let Wi1 ,i2 ,··· ,iv denote the system connection matrix of the resultant sys-
tem obtained by removing plants gi1 , gi2 , · · · , giv from the system. For each
gik , k = 1, 2, · · · , v, the removal of gik from the system is equivalent to the
system where there is no signal out of gik and no signal injected into it.
Then it is obvious that matrix Wi1 ,i2 ,··· ,iv is identical to the matrix formed
by deleting columns (i1 , i2 , · · · , iv ) and rows (i1 , i2 , · · · , iv ) from W . For ex-
ample, if plants g3 and g5 are removed from the system in Figure 3.5, then its
signal flow graph becomes Figure 3.6. It follows from (3.11) that the system
connection matrix is
 
1 0 0 −g1 −g1
 −g2 1 0 0 0 
 
W3,5 =   0 0 1 0 0 . (3.13)
 0 00 1 0 
−g7 0 0 0 1

g7 g4

g1 g2

g6

Fig. 3.6. Signal Flow Graph by deleting Plants g3 and g5 in Figure 3.5
74 3. Stability and Robustness of Feedback Systems

Indeed, (3.13) is identical to the matrix in (3.12) after deleting the third and
fifth columns and the third and fifth rows. Recall (Griffel, 1989) that a minor
of a matrix is the determinant of its submatrix. Let Mji11,j ,i2 ,··· ,iv
2 ,··· ,jv
be the minor
of W formed by deleting rows (i1 , i2 , · · · , iv ) and columns (j1 , j2 , · · · , jv ) from
W . We have just shown that matrix Wi1 ,i2 ,··· ,iv is identical to the matrix
formed by deleting columns (i1 , i2 , · · · , iv ) and rows (i1 , i2 , · · · , iv ) from W .
Thus, the following fact is true.
Fact 1 det(Wi1 ,i2 ,··· ,iv ) = Mii11,i,i22,···i
,··· ,iv
v
.
The key idea to prove Theorem 3.1.2 is to expand det(W ) step by step and
express it in terms of smaller sizes of minors until those minors are reached
which correspond to some closed-paths in the system signal graph. For later
reference, we define a closed-path as a signal flow path which starts from
a plant, may pass through other plants and nodes, and comes back to the
same plant. It should be noted that a closed-path may pass through a node
many times but a loop (Mason, 1956) is a special closed-path which can
pass through a node only once. For example, g1 , g2 , g3 and g6 form a loop,
and g1 , g2 , g5 and g7 form a closed-path but not a loop. Turn back to the
expansion of det(W ), and use W in (3.12) as an example. At the first stage,
we expand det(W ) by the first column as
ü ü
ü 1 0 0 0 0 −g1 −g1 üü
ü
ü −g2 1 0 0 −g2 0 0 üü
ü
ü 0 −g3 1 −g3 0 0 0 üü
ü
det(W ) = üü 0 0 −g4 1 0 0 0 üü
ü 0 −g5 0 −g5 1 0 0 ü
ü ü
ü 0 0 −g 0 0 1 0 ü
ü 6 ü
ü −g7 0 0 0 −g7 0 1 ü
ü ü
ü 0 0 0 0 −g1 −g1 üü
ü
ü −g3 1 −g3 0 0 0 üü
ü
ü 0 −g 1 0 0 0 üü
= det(W1 ) + (−1)4 × g2 üü 4
ü
ü −g 5 0 −g 5 1 0 0 ü
ü 0 −g6 0 0 1 0 ü
ü ü
ü 0 0 0 −g7 0 1 ü
ü ü
ü 0 0 0 0 −g1 −g1 üü
ü
ü 1 0 0 −g2 0 0 üü
ü
ü −g3 1 −g3 0 0 0 üü
ü
9 ü
+(−1) × g7 ü 0 −g4 1 0 0 0 üü
ü −g5 0 −g5 1 0 0 üü
ü
ü 0 −g6 0 0 1 0 üü
ü
ü ü
= det(W1 ) + (−1)4 g2 M12 + (−1)9 g7 M17 . (3.14)
One may note from (3.14) that det(W ) has two items left in addition to
det(W1 ). This is because the output of g1 is injected only into g2 and g7 . The
 3.1 Internal Stability 75

second stage of the expansion then deals with those nonzero minors (M12 and
M17 ) left from the first stage expansion. The columns on which we expand
these minors further are specific and are chosen as follows. For the minor M1i ,
we expand it on the i-th column of W (note that it is the i-th column of W
but not of M1i ). For M12 , this rule leads to
ü ü
ü 0 0 0 0 −g1 −g1 üü
ü
ü −g3 1 −g3 0 0 0 üü
ü
ü 0 −g 1 0 0 0 üü
M12 = üü 4
ü
−g
ü 5 0 −g 5 1 0 0 ü
ü 0 −g6 0 0 1 0 ü
ü ü
ü 0 0 0 −g7 0 1 ü
ü ü ü ü
ü 0 0 0 −g1 −g1 üü ü 0 0 0 −g1 −g1 üü
ü ü
ü −g4 1 0 0 0 üü ü 1 −g3 0 0 0 üü
ü ü
= (−1)4 g3 üü 0 −g5 1 0 0 üü + (−1)6 g5 üü −g4 1 0 0 0 üü
ü −g6 0 0 1 0 ü ü −g6 0 0 1 0 ü
ü ü ü ü
ü 0 0 −g7 0 1 ü ü 0 0 −g7 0 1 ü
2,3 2,5
= (−1)4 g3 M1,2 + (−1)6 g5 M1,2 , (3.15)

2,3 2,5
from which one may note once again that M1,2 and M1,2 are left because the
output of g2 is injected into g3 and g5 . Similarly, we expand M17 on the 7-th
column of W as
ü ü
ü 0 0 0 0 −g1 −g1 üü
ü
ü 1 0 0 −g2 0 0 üü
ü
ü −g 1 −g 0 0 0 üü
M17 = üü 3 3

ü 0 −g 4 1 0 0 0 üü
ü −g5 0 −g5 1 0 0 ü
ü ü
ü 0 −g6 0 0 1 0 ü
ü ü
ü 1 0 0 −g2 0 üü
ü
ü −g3 1 −g3 0 0 ü
ü ü
= (−1)8 g1 üü 0 −g4 1 0 0 üü = (−1)8 g1 M1,7
7,1
. (3.16)
ü −g5 0 −g5 1 0 ü
ü ü
ü 0 −g6 0 0 1ü
Note from Figure 3.5 that the output of g7 comes back to the input of g1
7,1
and form a closed-path. The corresponding minor M1,7 , by Fact 1, is equal
to det(W1,7 ), the determinant of the system connection matrix formed by
removing all plants in this closed-path. In general, if the plants involved
in a particular minor Mji11,j ,i2 ,··· ,iv
2 ,··· ,jv
form a closed-path, i.e., {i1 , i2 , · · · , iv } =
{j1 , j2 , · · · , jv } (namely, two sets are equal), we stop expanding this minor
and express it as det(Wi1 ,i2 ,··· ,iv ); Otherwise, continue to expand the minor.
With this rule, we proceed from (3.15) as
76 3. Stability and Robustness of Feedback Systems

ü ü
ü 0 0 0 −g1 −g1 üü
ü
ü −g41 0 0 0 üü
ü
2,3
M1,2 = üü 0 −g5 1 0 0 üü
ü −g60 0 1 0 üü
ü
ü 0 0 −g7 0 1 ü
ü ü ü ü
ü 0 0 −g1 −g1 üü ü 0 0 −g1 −g1 üü
ü ü
4 ü
ü −g 1 0 0 ü ü
6 ü 1 0 0 0 üü
= (−1) g4 ü 5 ü
0 0 1 0 ü + (−1) g6 ü −g5 1 0 0 üü
ü ü ü
ü 0 −g7 0 1 ü ü 0 −g7 0 1 ü
2,3,4 2,3,6
= (−1)4 g4 M1,2,3 + (−1)6 g6 M1,2,3 , (3.17)
and
ü ü
ü 0 0 0 −g1 −g1 üü ü ü
ü ü 0 0 −g1 −g1 üü
ü 1 −g3 0 0 0 ü ü ü
ü ü
9 ü 1 −g3 0 0 üü
2,5
M1,2 = üü −g4 1 0 0 ü
0 ü = (−1) g7 ü
ü −g6 ü −g4 1 0 0 üü
ü 0 0 1 0 üü ü −g6
ü 0 0 1 0 ü
0 −g7 0 1 ü
2,5,7
= (−1)9 g7 M1,2,5 . (3.18)
2,5,7 2,3,4 2,3,6
Since the minors M1,2,5 , M1,2,3 and M1,2,3 do not satisfy the condition on
stopping expansions, we continue to expand them as
ü ü
ü 0 0 −g1 −g1 üü ü ü
ü ü 1 −g3 0 ü
ü 1 −g3 0 ü
0 ü ü ü
2,5,7
M1,2,5 = üü ü = (−1)6 g1 üü −g4 1 0 üü
ü −g4 1 0 0 ü ü −g6 0 1 ü
ü −g6 0 1 0 ü
2,5,7,1
= (−1)6 g1 M1,2,5,7 = (−1)6 g1 det(W1,2,5,7 ), (3.19)
where it has been noticed that plant g1 , g2 , g5 and g7 form a closed-path,
which is clear from Figure 3.5 and also from the fact {1, 2, 5, 7} = {2, 5, 7, 1},
and no further expansion on W1,2,5,7 will be carried out. Similarly,
ÿ ÿ
ÿ 0 −g1 −g1 ÿ
ÿ ÿ
2,3,4
M1,2,3 = (−1) g5 ÿÿ 0
4
1 0 ÿÿ = (−1)4 g5 M1,2,3,4
2,3,4,5
,
ÿ −g7 0 1 ÿ
ÿ ÿ
ÿ −g −g1 ÿ
= (−1)4 g5 (−1)5 g7 ÿÿ 1 ÿ = (−1)9 g5 g7 M1,2,3,4,5
2,3,4,5,7
1 0 ÿ
= (−1)9 g5 g7 (−1)4 g1 M1,2,3,4,5,7
2,3,4,5,7,1
= (−1)13 g5 g7 g1 det(W1,2,3,4,5,7 ); (3.20)
ÿ ÿ
ÿ 1 0 0ÿ ÿ
ÿ
2,3,6
M1,2,3 = (−1)5 g1 ÿÿ −g5 1 0 ÿÿ = (−1)5 g1 M1,2,3,6
2,3,6,1
= (−1)5 g1 det(W1,2,3,6 ), (3.21)
ÿ 0 −g7 1 ÿ

where all the minors left have reached the corresponding closed-paths. To
be more clear, we substitute expressions (3.21), (3.20), (3.19), (3.18), (3.17),
(3.16), (3.15) into (3.14), and the expanding procedure looks as
 3.1 Internal Stability 77

det(W )
= det(W1 ) + (−1)4 g2 M12 + (−1)9 M17
2,3 2,5 7,1
= det(W1 ) + (−1)4 g2 [(−1)4 g3 M1,2 + (−1)6 g5 M1,2 ] + (−1)9 g7 (−1)8 g1 M1,7
2,3,4 2,3,6
= det(W1 ) + (−1)4 g2 {(−1)4 g3 [(−1)4 g4 M1,2,3 + (−1)6 g6 M1,2,3 ]
2,5,7
+(−1)6 g5 (−1)9 g7 M1,2,5 } + (−1)g1 g7 det(W1,7 )
2,3,4,5
= det(W1 ) + (−1)4 g2 {(−1)4 g3 [(−1)4 g4 (−1)4 g5 M1,2,3,4
2,3,6,1 2,5,7,1
+(−1)6 g6 (−1)5 g1 M1,2,3,6 ] + (−1)6 g5 (−1)9 g7 (−1)6 g1 M1,2,5,7 }
+(−1)g1 g7 det(W1,7 )
2,3,4,5,7
= det(W1 ) + (−1)4 g2 {(−1)4 g3 [(−1)4 g4 (−1)4 g5 (−1)5 g7 M1,2,3,4,5
+(−1)6 g6 (−1)5 g1 det(W1,2,3,6 )] + (−1)6 g5 (−1)9 g7 (−1)6 g1 det(W1,2,5,7 )}
+(−1)g1 g7 det(W1,7 )

= det(W1 )
+(−1)4 g2 {(−1)4 g3 [(−1)4 g4 (−1)4 g5 (−1)5 g7 (−1)4 g1 det(W1,2,3,4,5,7 )
+(−1)6 g6 (−1)5 g1 det(W1,2,3,6 )] + (−1)6 g5 (−1)9 g7 (−1)6 g1 det(W1,2,5,7 )}
+(−1)g1 g7 det(W1,7 )
= det(W1 )
+(−1)5 g2 g3 g4 g5 g7 g1 det(W1,2,3,4,5,7 ) + (−1)3 g2 g3 g6 g1 det(W1,2,3,6 )
+(−1)3 g2 g5 g7 g1 det(W1,2,5,7 ) + (−1)g7 g1 det(W1,7 ).

At a result, det(W ) has been expressed by the summation of det(W1 ) and

det(Wi1 ,i2 ,··· ,iv ), where plants (i1 , i2 , · · · , iv ) form closed-paths which pass
through plant g1 .
Let δi (1, · · · ), i = 1, 2, · · · , n1 , be all the distinct closed-paths each of which
passes through g1 , Fδi (1,··· ) the gain of the closed-path δi (1, · · · ), vδi the
total number of plants involved in the closed-path δi , and Wδi the system
connection matrix formed by removing all the plants involved in the closed-
path δi from the signal graph. With the above expanding procedure, we can
establish the following lemma for a general connection matrix W .

Lemma 3.1.1. Any system connection matrix W can be expanded as

n1
X
det(W ) = det(W1 ) − Fδi (1,··· ) det(Wδi (1,··· ) ). (3.22)
i=1
78 3. Stability and Robustness of Feedback Systems

Proof. A system connection matrix W has the form of

 
1 ∗ ··· ∗
∗ 
 
W = [wij ]n×n =  . ,
 .. W1 
∗

where, for i 6= 1, wi1 = gi if the output of plant g1 is one of the inputs of

plant gi ; otherwise, wi1 = 0. Expanding det(W ) by its first column gives

X i
det(W ) = det(W1 ) + (−1)pα1 +1 giα1 M1α1 , (3.23)
iα1

i
where the output of plant g1 is an input to plant giα1 . Next, we expand M1α1
by the iα1 -th column in W as
i
X iα1 ,iα2
M1α1 = (−1)pα2 +1 giα2 M1,i α
,
1
iα2

where the output of plant giα1 is an input to plant giα2 . If iα2 = 1, i.e., plant
α1 α2 i ,i
giα1 and plant g1 form a closed-path, then stop expanding M1,i α1
and we
have
i
α1 ,iα2 1,i
M1,i α
= M1,iαα1 = det(W1,iα1 ).
1 1

Compared with the Expansion Theorem (Griffel, 1989) of a determinant care-

fully, the term (−1)pα1 +pα2 is equal to (−1)γ , where γ = 1 is the minimal
number of permutations of {iα1 , 1} defined by Griffel (1989). Otherwise, for
iα1 ,iα2
those iα2 6= 1, expanding M1,i α
by the iα2 -th column in W yields
1

iα1 ,iα2
X i ,iα2 ,iα3
M1,i = (−1)pα3 +1 giα3 M1,i
α1
,
α1 α 1 ,iα2
iα3

where the output of plant giα2 is one of the inputs of plant giα3 . Again, if
iα3 = 1, i.e., plants giα1 , giα2 andg1 form a closed-path, then stop expanding
and we have
i
α1 ,iα2 ,iα3 1,i ,i
M1,i = M1,iαα1,iαα2 = det(W1,iα1 ,iα2 ),
α 1 ,iα2 1 2

where (−1)pα1 +pα2 +pα3 = (−1)γ , γ = 2 is the minimal number of permuta-

tions of {iα1 , iα2 , 1} defined by Griffel (1989). Otherwise, for those iα3 6= 1,
iα1 ,iα2 ,iα3
continue to expand M1,i α1 ,iα2
by the iα3 -th column in W . It follows from
this procedure that
 3.1 Internal Stability 79

det(W )
X iα1
= det(W1 ) + (−1)pα1 +1 giα1 M1
iα1
X X iα ,iα2
= det(W1 ) + (−1)pα1 +1 giα1 (−1)pα2 +1 giα2 M1,i1α
1
iα1 iα2

X
= det(W1 ) + (−1)pα1 +pα2 +2 Fδi det(Wδi )
iα2 =1,vδ =2
i
X pα1 +1
X iα ,iα2
+ (−1) giα1 (−1)pα2 +1 giα2 M1,i1α
1
iα1 iα2 6=1

X
= det(W1 ) + (−1)pα1 +pα2 +2 Fδi det(Wδi )
iα2 =1,vδ =2
i
X pα1 +1
X X iα ,iα2 ,iα3
+ (−1) giα1 (−1)pα2 +1 giα2 (−1)pα3 +1 giα3 M1,i1α
1 ,iα2
iα1 iα2 6=1 iα3

X
= det(W1 ) + (−1)pα1 +pα2 +2 Fδi det(Wδi )
iα2 =1,vδ =2
i

X
+ (−1)pα1 +pα2 +pα3 +3 Fδi det(Wδi )
iα3 =1,vδ =3
i
X X X iα ,iα2 ,iα3
+ (−1)pα1 +1 giα1 (−1)pα2 +1 giα2 (−1)pα3 +1 giα3 M1,i1α
1 ,iα2
iα1 iα2 6=1 iα3 6=1

···
X
= det(W1 ) + (−1)pα1 +pα2 +2 Fδi det(Wδi )
iα2 =1,vδ =2
i

X
+ (−1)pα1 +pα2 +pα3 +3 Fδi det(Wδi )
iα3 =1,vδ =3
i

X
+··· + (−1)pα1 +pα2 +···+pαx +x Fδi det(Wδi )
iαv̄ =1,vδ =v̄
i
X
= det(W1 ) + (−1)pα1 +pα2 +···+pαv +v Fδi det(W1,iα1 ,··· ,iαv−1 ),
vδ
i

where v̄ = max{vδi , i = 1, 2, · · · , n1 } and plants (iα1 , iα2 , · · · , iαv−1 , 1)

form a closed-path which passes through plant g1 . (−1)pα1 +pα2 +···+pαv =
(−1)γ , where γ = v − 1 is the minimal number of that permutations of
{iα1 , iα2 , · · · , iαv−1 , 1} (Griffel, 1989). Hence (3.22) is true. 2
We now turn to consider the system determinant ∆ and attempt to ex-
pand it into a similar form to (3.22). Let ρ(k1 , k2 , · · · , kiv ) denote a loop which
contains plants gi , i = k1 , k2 , · · · , kiv , and then F(k1 ,k2 ,··· ,kiv ) , gk1 gk2 · · · gkiv
is the gain of this loop. For ease of reference, we may also denote a loop by
80 3. Stability and Robustness of Feedback Systems

g4

g2
g3

g5

g6

¯ρ1
Fig. 3.7. Signal Flow Graph for ∆

ρi in case that the involved individual plants in the loop are clear from the
context. For a loop ρi , let ∆¯ρi denote the system determinant for the system
formed by removing loop ρi . Here, it should be pointed out that removing
a loop from the system means that not only all the plants but also all the
nodes in the loop are removed from the system. Let ∆ρi denote the system
determinant for the system formed by removing all the plants in loop ρi but
retaining all the relevant nodes. For example, ρ1 = ρ(1, 7) is a loop of the
system shown in Figure 3.5, then the graph formed by removing ρ1 is shown
in Figure 3.7 and the graph by removing all the plants in ρ1 only is shown
in Figure 3.8. Furthermore, ∆¯ρ1 = 1 − g3 g4 and ∆ρ1 = 1 − g2 g5 − g3 g4 . In
general, ∆¯ρi 6= ∆ρi .

g7 g4

g1 g2 g3

g5

g6

Fig. 3.8. Signal Flow Graph for ∆ρ1

If plant g1 is removed from a given system with its system determinant

∆, the system determinant of the resultant system is denoted by ∆1 . Then, it
follows from (3.8) that ∆ may be decomposed into two parts: one is indepen-
dent of g1 and the system determinant ∆1 of the system formed by removing
plant g1 ; and other is related to plant g1 , i.e. ∆ − ∆1 = −g1 Z for some Z.
For example, consider the system determinant shown in Figure 3.5:
 3.1 Internal Stability 81

X X
∆ = 1− F1i + F2j = 1 − g1 g7 − g2 g5 − g3 g4 − g1 g2 g3 g6 + g1 g7 g3 g4
i j
= (1 − g2 g5 − g3 g4 ) − g1 g2 g3 g6 − g1 g7 + g1 g7 g3 g4 . (3.24)

The signal flow graph after removing g1 is shown in Figure 3.9 and its system
determinant is ∆1 = 1 − g3 g4 − g2 g5 , which is, of course, independent of g1 .
The remaining part which is related to plant g1 is

∆ − ∆1 = −g1 g2 g3 g6 − g1 g7 + g1 g7 g3 g4 = −g1 Z

for Z = g2 g3 g6 + g7 − g7 g3 g4 . Look into each item in g1 Z and one sees that

it is related to one and only one loop containing g1 , which is Fρ1 = g1 g7 or
Fρ2 = g1 g2 g3 g6 . More precisely, g1 Z can be written as g1 Z = g1 g2 g3 g6 (1) +
g1 g7 (1 − g3 g4 ) = Fρ(1,2,3,6) ∆¯ρ(1,2,3,6) + Fρ(1,7) ∆¯ρ(1,7) . For a general system, g1 Z
is the sum of loop gains and the products of those nontouching loop gains for
those loops containing g1 , and in all these items the plant g1 appears once
and only once. This fact is due to Mason (1956).

g7 g4

g2
g3

g5

g6

Fig. 3.9. Signal Flow Graph corresponding to ∆1

P
Lemma 3.1.2. ∆ = ∆1 − Fρi (1,··· ) ∆¯ρi (1,··· ) , where ρi (1, · · · ) are the
i
loops each of which contains plant g1 .
Now, we consider the relationship between ∆ρi and ∆¯ρi . Assume that loop
ρj has a common node with ρi , then Fρj will not appear in ∆¯ρi . It is because
this node has been removed from the graph when we calculate ∆¯ρi and the
relevant plants involved in the original ρj does not form a loop in the new
system anymore. However, Fρj will appear in ∆ρi because only the plants
involved in ρi are removed from the original graph when we calculate ∆ρi ,
but the relevant nodes are retained so that ρj is still a loop in the resultant
graph. For instance, consider the system in Figure 3.5. For ρi = ρ(1, 7), we
have ∆ρi = 1 − g2 g5 − g3 g4 and ∆¯ρi = 1 − g3 g4 . The difference between them
is g2 g5 , the gain of the ρj = ρ(2, 5) which is touched with loop ρi only by
82 3. Stability and Robustness of Feedback Systems

one node. It follows that ∆¯ρi − ∆ρi = g2 g5 = Fρ(2,5) . One further sees that
Fρi (∆¯ρi − ∆ρi ) = Fρi Fρj is the gain of the closed-path which is formed by
ρ(1, 7) and ρ(2, 5) that have a common node. For a general system, let ρij
(ρi,j
k , · · · , respectively) be the loops each of which has a common node with
ρi(1,··· ) (ρi or ρj ,· · · , respectively) but does not contain the plant g1 . It follows
from Lemma 3.1.2 that
X
∆ = ∆1 − Fρ (1,··· ) ∆¯ρ (1,··· )
i i
i
X
= ∆1 − Fρi (1,··· ) (∆ρi (1,··· ) + ∆¯ρi (1,··· ) − ∆ρi (1,··· ) )
i
X X
= ∆1 − Fρi (1,··· ) ∆ρi (1,··· ) − Fρi (1,··· ) (∆¯ρi (1,··· ) − ∆ρi (1,··· ) )
i i
X X X
= ∆1 − Fρi (1,··· ) ∆ρi (1,··· ) − Fρi [ Fρij ∆¯ρi ρij
i i j
X
− Fρij Fρij ∆¯ρi ρij ρij + ···]
1 2 1 2
j1 ,j2
X X X
= ∆1 − Fρi (1,··· ) ∆ρi (1,··· ) − Fρi Fρij ∆ρi ρij
i i j
X X X
− Fρi [ Fρij (∆¯ρi ρij − ∆ρi ρij ) − Fρij Fρij ∆¯ρi ρij ρij + · · ·]
1 2 1 2
i j j1 ,j2
X X X
= ∆1 − Fρi (1,··· ) ∆ρi (1,··· ) − Fρi Fρij ∆ρi ρij
i i j
X X
− Fρi ( Fρij Fρi,j ∆¯ρi ρi ρi,j + · · ·)
k j k
i j,k
= ··· ,
where loops ρij1 andρij2 are nontouched to each other but both touched with
loop ρi , ρi,j i
k the loop touched with loops ρi or ρj , ∆ρi ρi ρi,j ··· the system deter- j k

minant by removing all the plants in loops ρi , ρij , ρi,j k , · · · from the system. It
is noted that the set of the closed-paths δi (1, · · · ) each of which passes through
g1 consists of the loops ρi (1, · · · ) and the non-loop closed-paths formed by
ρi (1, · · · ) plus ρij , ρi (1, · · · ) plus ρij and ρi,j
k , · · · . Thus, we have the following
lemma.
P
Lemma 3.1.3. ∆ = ∆1 − Fδi (1,··· ) ∆δi (1,··· ) .
i

For example, there exist four closed-paths through plant g1 in Figure 3.5
and they are δ1 (1, 7), δ2 (1, 2, 5, 7), δ3 (1, 2, 3, 4, 5, 7) and δ4 (1, 2, 3, 6). The cor-
responding gains are Fδ1 = g1 g7 , Fδ2 = g1 g2 g5 g7 , Fδ3 = g1 g2 g3 g4 g5 g7 and
Fδ4 = g1 g2 g3 g6 . Obviously, δ1 = ρ1 and δ4 = ρ2 are loops, but δ2 and δ3 are
not. We have ∆1 = 1 − g2 g5 − g3 g4 , ∆δ1 = 1 − g3 g4 − g2 g5 , ∆δ2 = 1 − g3 g4 ,
∆δ3 = 1 and ∆δ4 = 1. It then follows that
 3.1 Internal Stability 83

∆ 1 − Fδ 1 ∆ δ 1 − Fδ 2 ∆ δ 2 − Fδ 3 ∆ δ 3 − Fδ 4 ∆ δ 4
= (1 − g3 g4 − g2 g5 ) − g1 g7 (1 − g3 g4 − g2 g5 ) − g1 g2 g5 g7 (1 − g3 g4 )
− g1 g2 g3 g4 g5 g7 (1) − g1 g2 g3 g6 (1)
= 1 − g3 g4 − g2 g5 − g1 g7 − g1 g2 g3 g6 + g1 g7 g3 g4 ,

which, by (3.24), is equal to ∆, indeed. It may also be verified that

Fρ (∆¯ρ − ∆ρ ) = (g1 g7 )(g2 g5 ) = g1 g7 g2 g5 (1 − g3 g4 + g3 g4 )
1 1 1

= g1 g7 g2 g5 (1 − g3 g4 ) + g1 g7 g2 g5 g3 g4 (1)
= Fδ 2 ∆ δ 2 + F δ 3 ∆ δ 3 ,
and
Fρ2 (∆¯ρ2 − ∆ρ2 ) = g1 g2 g3 g6 (1 − 1) = 0,
so that

∆1 − Fρ1 ∆¯ρ1 − Fρ2 ∆¯ρ2

= ∆1 − Fρ1 ∆ρ1 − Fρ2 ∆ρ2 − Fρ1 (∆¯ρ1 − ∆ρ1 ) − Fρ2 (∆¯ρ2 − ∆ρ2 )
= ∆ 1 − Fδ 1 ∆ δ 1 − Fδ 2 ∆ δ 2 − Fδ 3 ∆ δ 3 − Fδ 4 ∆ δ 4 .

With Lemmas 3.1.1 and 3.1.3, we are now in the position to prove Theo-
rem 3.1.2.

Proof of Theorem 3.1.2 By Theorem 3.1.1, the proof will be completed

if there holds

∆ = det(W ). (3.25)

We will prove (3.25) by induction. If the system has only one plant, then
det(W ) = 1 and ∆ = 1, (3.25) is thus true. When the system contains two
plants g1 and g2 , then n = 2 and there are only two cases to be considered.
Case 1 is that these two plants form a loop, and then its connection matrix
is
 
1 −g1
W = ,
−g2 1
and det(W ) = 1 − g1 g2 . From (3.8), ∆ = 1 − g1 g2 . (3.25) thus holds in this
case. Case 2 is that these two plants do not form a loop, and then we have
 
10
W = ,
01

and det(W ) = ∆ = 1. (3.25) is thus proven for the case of two plants. Assume
now that (3.25) holds for systems with (n − 1) plants, we need to show that
it is also true for the system with n plants. The above assumption implies
that det(W1 ) = ∆1 for (n − 1) plants and det(Wδi ) = ∆δi for (n − 1) plants.
Substituting them into (3.22) yields
84 3. Stability and Robustness of Feedback Systems

X
det(W ) = ∆1 − Fδi(1,··· ) ∆δi(1,··· )
i
= ∆,
where the last equality follows from Lemma 3.1.3. Hence (3.25) holds true
for any integer n. 2
In this section, two theorems on internal stability of interconnected sys-
tems, one for scalar signals and other for vector signals, have been established.
They are stated in terms of a single polynomial only. This polynomial is read-
ily found from a given system and can be read off by inspection in many cases.
The application to a few typical control systems has shown simplicity and
effectiveness of the results.

3.2 Unity Feedback Systems

A typical feedback control system has a plant, a sensor, a controller K(s), and
an actuator as its principal components. The control objective is to make the
plant output to track the reference (or set-point) R(s) as closely as possible
in the presence of possible disturbances V (s). For system analysis and design,
the plant, sensor, and actuator are usually combined and viewed as a single
object (still called the plant G(s)), and this gives rise to the most popular
control configuration, the unity (output) feedback system, as depicted in
Figure 3.10. We assume throughout the rest of this chapter that G(s) and
K(s) are proper and there holds det(I + G(∞)K(∞)) 6= 0. The system can
be described by the input-output relationship:
   
E(s) R(s)
= H(s) .
U (s) V (s)
To get H(s), it follows from inspection of Figure 3.10 that
E(s) + G(s)U (s) = R(s), (3.26)
−K(s)E(s) + U (s) = V (s), (3.27)
leading via the matrix inversion to
 −1
I G(s)
H= , (3.28)
−K(s) I
and via direct variable elimination to
E(s) + G(s)(V (s) + K(s)E(s)) = R(s),
−K(s)(R(s) − G(s)U (s)) + U (s) = V (s),
or
 
(I + GK)−1 −(I + GK)−1 G
H= . (3.29)
(I + KG)−1 K (I + KG)−1
 3.2 Unity Feedback Systems 85

V (s)
+
R (s) + E (s) U (s ) Y (s )
K (s) G (s)
_ +

Fig. 3.10. Unity Feedback System

−1
Let G = D−1 N = NR DR be coprime PMFs of G and K = Y X −1 =
XL−1 YLbe coprime PMFs of K, pG and pK be the characteristic polynomials
of G and K, respectively.
Theorem 3.2.1. The following are equivalent:
(i) The system in Figure 3.10 is internally stable;
(ii) pc := pG pK det[I + GK] = pG pK det[I + KG] is a stable polynomial;
(iii) det[DX + N Y ] is a stable polynomial, or det[XL DR + YL NR ] is a
stable polynomial;
(iv) H(s) is stable.
Proof. (i) ⇐⇒ (ii) It follows from Example 3.1.1 for the case of n = 2.

(ii) ⇐⇒ (iii) Notice det[DX + N Y ] = det(D) det[I + GK] det(X), and

it is shown in Chapter 2 that det(D) and det(X) are equal to pG and pK
modulo a nonzero constant factor, respectively. Hence det[DX +N Y ] is equal
to pc (s) modulo a nonzero constant, and so is det[XL DR +YL NR ] by a similar
reasoning.

(iii) ⇐⇒ (iv) Rewrite H in (3.28):

 −1
I D−1 N
H=
−Y X −1 I
 −1     −1 −1
D 0 DX N X 0
=
0 I −Y I 0 I
  −1  
X0 DX N D0
= , (3.30)
0 I −Y I 0 I
which is a left and right PMF of H(s). For the composite matrix,
 
DX N
 −Y I 
 
 X 0 ,
0 I
86 3. Stability and Robustness of Feedback Systems

subtract the last block row from the second block row to get
 
DX N
 −Y 0 
 
 X 0 ,
0 I

which is ensured to have full column rank due to the assumed coprimeness
of X and Y . Thus,
  −1
X0 DX N
0 I −Y I

is right coprime. Similarly, one can easily show that

 −1  
DX N D0
−Y I 0 I

is left coprime. Then, the pole polynomial for H is (see Chapter 2)

 
DX N
pH (s) = det .
−Y I

Notice
     
DX N DX N I 0
det = det det
−Y I −Y I Y I
  
DX N I 0
= det
−Y I Y I
 
DX + N Y N
= det
0 I
= det[DX + N Y ].

Hence stability of H is equivalent to that of the polynomial det[DX + N Y ],

and the proof is completed. 

The fact that stable det[DX + N Y ] implies stability of H can also be seen
in an explicit way as follows. By the following well-known matrix identities,

(I + KG)−1 K = K(I + GK)−1 ,

and
(I + KG)−1 = I − (I + GK)−1 KG
= I − K(I + GK)−1 G,

(3.29) may now be rewritten as

 3.2 Unity Feedback Systems 87

 
(I + GK)−1 −(I + GK)−1 G
H(s) = , (3.31)
K(I + GK)−1 I − K(I + GK)−1 G

which involves (I +GK)−1 only but not (I +KG)−1 . Substituting G = D−1 N

and K = Y X −1 into (3.31) yields
 
X(DX + N Y )−1 D −X(DX + N Y )−1 N
H(s) = .
Y (DX + N Y )−1 D I − Y (DX + N Y )−1 N

Obviously, stability of the polynomial det[DX + N Y ] implies stability of the

rational matrix H(s).
The important message conveyed by Theorem 3.3 (iv) is that internal
stability of the system is equivalent to (input-output) stability of H(s). But
note that all four blocks, Hij , i, j = 1, 2, in H, must be stable to ensure inter-
nal stability of the feedback system, and stability of all Hij can rule out the
possibility of unstable pole-zero cancellation between G and K. Otherwise,
unstable pole-zero cancellation may occur in the loop, and the system is not
internally stable. Yet, some (but not all) of Hij is stable. This is precisely
the problem associated with input-output stability, but can be revealed by
internal stability analysis.
Example 3.2.1. Let
" 1
#
0 s+1
G(s) = s 2
,
− s+2 s+2

"1 #
s 0
K(s) = 1
.
1 s

We find coprime fractions for G(s) and K(s):

 −1  
s+1 0 0 1
G(s) = ,
0 s+2 −s 2

  −1
10 s0
K(s) = .
s1 os

Then, pG (s) = (s + 1)(s + 2), pK (s) = s2 , and

" 1 1 #
(1 + s+1 ) s(s+1)
det[I + GK] = det 1 2
s+2 (1 + s(s+2) )
2
= ( s+2 s +2s+2
s+1 )( s(s+2) ) −
1
s(s+1)(s+2)

s3 +4s2 +6s+3
= s(s+1)(s+2) .
88 3. Stability and Robustness of Feedback Systems

We have
pc (s) = s(s3 + 4s2 + 6s + 3),
a unstable polynomial, so the system is unstable. One can easily verify that
det[DX + N Y ] = pc (s). One also find
" 1 1 #
s+1 s(s+1)
GK = 1 2
,
s+2 s(s+2)

H21 = (I + KG)−1 K
 3 2 
s +3s +4s+3 s+2
s(s3 +4s2 +6s+3) − s(s3 +4s 2 +6s+3)
= ,
s3 +4s2 +7s+4 1
s3 +4s2 +6s+3 s

which is unstable. ♦
Instability of this example is caused by some closed right half plane (RHP)
pole-zero cancellation between G and K. Indeed,
 −1   −1
s+1 0 s1 s0
GK =
0 s+2 s2 0s
 −1   −1
s+1 0 11 10
=
0 s+2 12 0s
indicates a RHP pole-zero cancellation at s = 0 between G and K.
Another cause for instability is that some RHP pole is lost when forming
det[I + GK], as can be seen in the following example.
Example 3.2.2. Let
" 1 #  
s+1 0 10
G(s) = and K= .
1 1 01
s−1 s+1

GK has no pole-zero cancellation as K has neither zero nor pole. Further,

one sees that
 2  2
1 s+2
det[I + GK] = 1 + =
s+1 s+1
has stable roots only. However, simple calculations give
s+2 2
pc (s) = pG pK det[I + GK] = (s − 1)(s + 1)2 · 1 · ( ) = (s − 1)(s + 2)2 ,
s+1

a unstable polynomial. Thus, the system is not internally stable. Instability

of the system can be clearly exhibited with Figure 3.11, where the unstable
1
element of the plant, s−1 , stays outside any loop and is impossible to be
stabilized by any feedback. ♦
 3.2 Unity Feedback Systems 89

_
u1 y1
1
1
s +1

1
s- 1

u2 + y2
1
1
_ s +1 +

Fig. 3.11. An unstable system

If K(s) and G(s) are both unstable, then it is necessary to check all four
blocks of the H-matrix to ensure internal stability. However, if at least one
of them is stable, then there is less work to do.

Theorem 3.2.2. If K(s) is stable, then the feedback system shown in Figure
3.10 is internally stable if and only if H12 (s) = −[I + G(s)K(s)]−1 G(s) is
stable.

Proof. “only if”: Since all Hij are stable, H12 is stable, too.
“if”: It follows from (3.29) that

H11 = (I + GK)−1
= (I + GK)−1 (I + GK − GK)
= I + H12 K

is stable since I, H12 and K are so. By (3.31), H21 = K(I + GK)−1 = KH11
is stable, and H22 = I − K(I + GK)−1 G = I − KH12 is stable, too. 

Theorem 3.2.1 leads us to examine the following polynomial matrix equa-

tion,

D(s)X(s) + N (s)Y (s) = P (s), (3.32)

where D(s) ∈ Rp×p [s], N (s) ∈ Rp×m [s], and P (s) ∈ Rp×p [s] are given (with
P (s) being stable in contents of control engineering), and X(s) ∈ Rp×p [s]
and Y (s) ∈ Rm×p [s] are sought to satisfy the equation. Such an equation is
called Diophantine equation. Equation (3.32) is linear in X and Y and this
enables us to establish the following results.
90 3. Stability and Robustness of Feedback Systems

Theorem 3.2.3. Consider (3.32).

(i) There is a solution pair of X(s) and Y (s) if and only if a greatest
common left divisor (GCLD) is a left divisor of P (s);
(ii) Let [D N] have full row rank, N bD b −1 be a right coprime PMF of
−1
D N , and (X0 , Y0 ) be a particular solution of (3.32), then the general so-
lution of (3.32) is given by

X(s) = X0 (s) + N b (s)T (s), (3.33a)

b
Y (s) = Y0 (s) − D(s)T (s), (3.33b)
for an arbitrary polynomial matrix T (s); and
(iii) If D(s) and N (s) are left coprime, then (3.32) always admits a so-
lution and has the general solution in the form of (3.33).

Proof. (i) Let L(s) be a GCLD of D(s) and N (s). Then, it follows that there
is a unimodular U (s) such that

[D N ]U = [L 0]

whose first block column gives

DU1 + N U2 = L (3.34)

for some polynomial matrices U1 and U2 . If L(s) is also a left divisor of P (s),
i.e. P (s) = L(s)P (s) for some polynomial matrix P , then post-multiplying
(3.34) by P yields

D(U1 P ) + N (U2 P ) = LP = P (s),

showing that X(s) = U1 P and Y (s) = U2 P meets (3.32), and thus form a
solution pair. Conversely, if (3.32) has a solution (X, Y ), it can be written as

LDX + LN Y = P,

or

L(DX + N Y ) = P,

so that L is a left divisor of P .

(ii) Since (X0 , Y0 ) is a solution to (3.32), there holds

DX0 + N Y0 = P. (3.35)

Subtracting (3.35) from (3.32) gives

D(X − X0 ) + N (Y − Y0 ) = 0,

or
 3.2 Unity Feedback Systems 91

 
  X − X0
DN = 0. (3.36)
Y − Y0
Since [D N] is of size p × (p + m) and rank p, the right null space of [D N]
is spanned by a (p + m) × m matrix B(s) such that [D N]B = 0 and B has
rank m. The general solution of the homogeneous equation (3.36) can then
be expressed as
 
X − X0
= BT (3.37)
Y − Y0

bD
for an arbitrary polynomial matrix T . From D−1 N = N b −1 , we have
" #
  N b
D N b = 0.
−D

b and D,
And, due to the assumed coprimeness of N b [N
bT − D
b T ]T has full rank
m. Thus, we can take
" #
b
N
B= b ,
−D

and substituting it into (3.37) produces the required form (3.33).

(iii) The identity matrix I is a GCLD of coprime D and N , and it is a left
divisor of any matrix like P (s). Hence, the result follows from (i) and (ii). 

Among all solutions of (3.32), there must be at least one, denoted by

(Xmin , Ymin ) such that it has minimum column degrees, i.e.
   
Xmin X
∂i ≤ ∂ci
Ymin Y
for any other solution (X, Y ). In view of (3.33), we can obtain it as follows.
Take column i of any solution
 
X
,
Y
and subtract from it a polynomial combination of the columns of
" #
Nb
−Db

so as to reduce its degree as much as possible. Repeat the process for all other
columns and the result would be (Xmin , Ymin ).
With obvious changes, we may work with the case where (X, Y ) is given
and (D, N ) is sought and get the dual results to Theorem 3.2.3.
Theorem 3.2.3 lays a foundation for stabilization and arbitrary pole as-
signment for any plant G(s). One may get its left coprime fraction D−1 (s)N (s)
92 3. Stability and Robustness of Feedback Systems

and specify a suitable P (s) which reflects the designer’s control specifications.
Then, (3.32) is solved to get a solution (X, Y ) and a stabilizer for G(s) is
obtained as K(s) = Y X −1 provided that X is nonsingular and Y X −1 proper.
This procedure actually does more than stabilization or pole assignment be-
cause it can assign an entire polynomial matrix P (s) instead of just det P (s),
the characteristic polynomial of the feedback system.

Example 3.2.3. Consider

1 
s 1 0
G= .
1
00 s
Suppose that a controller in the unity feedback system is to be found such
that the resulting closed-loop denominator is
 
(s + 1)2 0
P (s) = .
0 s+1
We write
 −1  
s0 1s0
G= ,
0s 001
which is left coprime. Take K = K1 K2 with
 
10
K1 =  0 0  .
01
Then, it follows that
 −1
s0
GK1 = I2
0s
is decoupled. Let
 b    −1
0 b0 (s + a) 0
K2 = s+a = .
0 c 0c 0 1
Then, (3.32) becomes
      
s0 s+a 0 b0 (s + 1)2 0
+ = ,
0s 0 1 0c 0 s+1
whose solution is a = 2, b = 1; and c = 1 so that the controller is
  1   1 
10 s+2 0 (s+1) 0
K = K1 K2 =  0 0    =  0 0,
01 0 1 0 1
which is also proper. ♦
 3.2 Unity Feedback Systems 93

In general, a solution (X, Y ) from Diophantine equation (3.32) may not

guarantee properness of Y X −1 . One has to pay a special attention to this
problem. Besides, for practical use, systematic numerical algorithms would
be desirable, especially, to be included in control system CAD toolbox. We
now present such an algorithm.
Given a p×m proper plant G(s), we obtain its left right fraction D−1 N =
G with D row reduced. Let νi , i = 1, 2, . . . , p, be the row degrees of D(s) and
ν = max{ν1 , ν2 , . . . , νp }. Let µ be the largest controllability index of G(s).
We write
D(s) = D0 + D1 s + . . . + Dν sν ,
N (s) = N0 + N1 s + . . . + Nν sν ,
X(s) = X0 + X1 s + . . . + Xq xq ,
Y (s) = Y0 + Y1 s + . . . + Yq xq ,
P (s) = P0 + P1 s + . . . + Pν+q sν+q .
Substituting these into (3.32) yields
 
D 0 N0 0 0 ··· 0 0
 .. ..   
P0

 D 1 N1 D 0 N0 · · · . .  D0
 
 .. .. .. .. . . .. ..   N0  
  P1 

 . . . . . . . 
  
  D1   .  .

 Dν Nν Dν−1 Nν−1 · · · 0
 0 
  N1   . 
 0 0 D ν Nν · · · D 0 N0  .  =  . . (3.38)
  .   .. 
 .. ..  .  

 . 

 0 0 
 0 0 ··· . .  Dν   .. 
 . . .. .. 
 .. .. . . ··· D N
ν−1 ν−1
 Nν Pν+q
0 0 0 0 · · · Dν Nν

Theorem 3.2.4. Choose an integer q such that q ≥ µ for proper G(s) or

q ≥ µ − 1 for strictly proper G(s). Specify the closed-loop denominator ma-
trix P (s) to have the desired characteristic polynomial pc (s) = det P (s) and
characteristic structure, and to meet the condition that
lim diag {s−ν1 , s−ν2 , . . . , s−νp }P (s)diag {s−q , s−q , . . . , s−q }
s→∞

exists and is nonsingular. Then, there exists a proper controller K(s) =

Y (s)X −1 (s) such that (3.32) holds true.
Proof. See Chen (1984). 
Quite often, one needs to investigate stability (we shall drop “internal”
from now on) as well as its margins when the loop gain varies. In this regard,
one may view G(s) ∈ Rm×m (s) in Figure 3.10 as the series connection of the
plant with the dynamic part of the controller, and we assume that there are no
RHP pole-zero cancellations between these two components. Let K(s) = kIm
for a real number k to examine system properties for various values of k.
Recall from Theorem 3.2.1 that
94 3. Stability and Robustness of Feedback Systems

pc (s) = pG (s)pK (s) det[I + G(s)K(s)],

where pc (s) is the pole or the characteristic polynomial of the closed-loop
system, pG (s) and pK (s) are the pole polynomials of G(s) and K(s) respec-
tively. Obviously, our present case leads to pK (s) = 1 for K(s) = kI and
pc (s) = pG (s) det[I + kG(s)], or
pc (s)
det[I + kG(s)] = .
pG (s)
The principle of the argument is applied to get
∆ arg det(I + kG(s)) = 2π(Pc − Po ),
where Pc and Po are the numbers of RHP roots of pc (s) and pG (s), respec-
tively, where pG (s) is the same as the open-loop characteristic polynomial in
the present case of static controller K(s) = K, so the notation Po is justified.
∆ arg denotes the change in the argument as s transverses the Nyquist con-
tour once. As usual, the Nyquist contour consists of the straight-line segment
from (0, −jR) to (0, jR) and the RHP semicircle joining these points, and by
convention it is traversed clockwise. If G(s) has any poles on the imaginary
axis, the contour has to be indented into the left half-plane with a small
semicircle surrounding each of them. For the closed-loop stability, we need
Pc = 0, i.e.
∆ arg det(I + kG(s)) = −2πPo .
Theorem 3.2.5. The feedback system in Figure 3.10 is stable if and only if
the Nyquist plot of det[I +kG(s)] encircles the origin Po times anti-clockwise.
To apply the above theorem, however, one has to draw the Nyquist plot
of det[I + kG(s)] for each value of k which is under consideration. This is
inconvenient, and different from the SISO case where one can know stability
for all values of k once the Nyquist plot of G(s) is drawn.
Let λi (s) be an eigenvalue of G(s). Then, kλi (s) is an eigenvalue of kG(s)
and (1 + kλi (s)) is that of (I + kG(s)). One gets
Y
det[I + kG(s)] = [1 + kλi (s)],
i

and
X
∆ arg det(I + kG(s)) = ∆ arg[1 + kλi (s)]. (3.39)
i

The Nyquist plot of λi (s) are called the characteristic loci of G(s). They
have m branches for an m × m plant G(s). It follows from (3.39) that we can
infer closed-loop stability by counting the total number of encirclements of
the point (−1 + j0) made by the characteristic loci kλi (jw) of kG(s), which
is as powerful as in the classical SISO Nyquist criterion as it is now enough
to draw λi (jw) once, usually for k = 1.
 3.2 Unity Feedback Systems 95

Theorem 3.2.6. (Multivariable Nyquist Theorem) Suppose that G(s)

has Po unstable poles. Then the feedback system in Figure 3.10 with K(s) =
kI is stable if and only if the characteristic loci of kG(s), taken together,
encircle the point, −1 + j0, Po times anti-clockwise.

Example 3.2.4. Let

 
0.5 2s + 1 1
G(s) = .
(s + 1)(s + 2) 1 2s + 3
Its characteristic loci are found to be
1
λ1 (jw) = jw+2 ,
1
λ2 (jw) = jw+1 ,

which are shown in Figure 3.12. The open loop is stable, i.e. Po = 0. For the
closed-loop system to be stable, there should be zero net encirclements of the
critical point (− k1 + j0) made by the loci. Thus, one infers that
• for −∞ < −1 k < 0, there is no encirclement and the closed-loop is stable;
• for 0 < −1
k < 1, there are one or two clockwise encirclements, implying
closed-loop instability; and
• for −1
k > 1, there is no encirclement, and closed-loop stability is ensured.
♦

Im
0.5

0.4

0.3

0.2

0.1
0.5 Re
0

0. 1

0. 2

0. 3

0. 4

0. 5

0. 2 0 0.2 0.4 0.6 0.8 1 1.2

Fig. 3.12. Characteristic loci

( —— λ1 , - - - λ2 )

It is noted that the computation of characteristic loci is quite involved in

general. For the special case of diagonally dominant systems, however, this
is not necessary, and only the so-called Gershgorin bands need to be drawn.
96 3. Stability and Robustness of Feedback Systems

Theorem 3.2.7. (Gershgorin’s Theorem): Let Z be a complex matrix of

m × m size, then the eigenvalues of Z lie in the union of the m circles, each
with centre zii and row radius:
m
X
|zij |, i = 1, 2, . . . , m.
j=1,j6=i

They also lie in the union of circles, each with centre zii and column radius:
m
X
|zji |, i = 1, 2, . . . , m.
j=1,j6=i

Proof. Let λ be an eigenvalue of Z and v = (v1 , v2 , · · · , vm )T the correspond-

ing eigenvector, then

Zv = λv, (totally m rows).

Consider only the i − th row for which |vi | ≥ |vj | for all j 6= i,
m
X
zij vj = λvi ,
j=1

or
m
X
λvi − zii vi = zij vj .
j=1,j6=i

It follows that
m
X m
X m
X
vj vj
|λ − zii | = | zij |≤ |zij | · | |≤ |zij |,
vi vi
j=1,j6=i j=1,j6=i j=1,j6=i

hence, the result. 

Consider now a proper rational matrix G(s). Its elements are functions
of a complex variable s. The Nyquist array of G(s) is an array of graphs, the
(i, j) graph of which is the Nyquist plot of gij (s). For any i, at each s = jw,
a circle of either row radius:
m
X
|gij (jw)|,
j=1,j6=i

or column radius
m
X
|gji (jw)|,
j=1,j6=i
 3.2 Unity Feedback Systems 97

is superimposed on gii (jw). The “bands” obtained in this way are called the
Gershgorin bands. For illustration, Figure 3.13 shows the Nyquist array and
Gershgorin bands of
" s+2 s+1 #
2s+1 2s+1
G(s) = s+1 s+2
. (3.40)
2s+1 2s+1

1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5

−2 −2
−1 0 1 2 3 −1 0 1 2 3
element (1,1) element (1,2)

1.5 1.5

1 1

0.5 0.5

0 0

−0.5 −0.5

−1 −1

−1.5 −1.5

−2 −2
−1 0 1 2 3 −1 0 1 2 3
element (2,1) element (2,2)

Fig. 3.13. Nyquist Array with Gershgorin Bands

By the Gershgorin’s theorem, for a given s = jw, the eigenvalues (char-

acteristic loci) λ(jw) lie in the Gershgorin circles of G(jw). When s traverses
the Nyquist contour, the union of the Gershgorin bands ”traps” the union of
the characteristic loci. If all the Gershgorin bands exclude the point (1 + j0),
then it is sufficient to infer closed-loop stability by counting the encircles of
(1 + j0) by the Gershgorin bands, since in this case, these encirclements are
the same as those by the characteristic loci. In this case, [I + G] is diagonal
dominant, a concept which we now introduces formally.
98 3. Stability and Robustness of Feedback Systems

Definition 3.2.1. (Diagonal Dominance) A m × m proper rational matrix

G(s) is called row diagonally dominant (on the Nyquist contour) if
m
X
|gii (jw)| > |gij (jw)|, i = 1, 2, . . . , m;
j=1,j6=i

and column diagonally dominant if

m
X
|gii (jw)| > |gji (jw)|, i = 1, 2, . . . , m,
j=1,j6=i

assuming that G(s) has no poles on the Nyquist contour.

Graphical Test for Dominance can be easily developed. Draw the Gersh-
gorin bands for G(s). If the banks with row radius exclude the origin, then
G(s) is row diagonally dominant. If it is the case with column radius, G(s)
is column diagonally dominant. A greater degree of dominance corresponds
to narrower Gershgorin bands and more closely resembles m decoupled SISO
systems. For G(s) in (3.40), Figure 3.13 indicates that G(s) is diagonally
dominant, which can be confirmed analytically by noting |jw + 2| > |jw + 1|
for any real w.
For diagonally dominant G(s), we have the following stability criterion.

Theorem 3.2.8. (Diagonally dominant Systems) Consider the unity

feedback system with the loop gain L(s) = G(s)K, where G(s) is square with
Po unstable poles (multiplicities included), K = diag {ki }, and
1 X
|gii (s) + |> |gji (s)|,
ki
j6=i

for each i and for all s on the Nyquist contour, that is I + GK is column
diagonally dominant. Let the Nyquist plot of gii (s) encircle the point, (−1/ki +
j0), Ni times anticlockwise. Then, the system is stable if and only if
m
X
Ni = Po .
i=1

Proof. Due to the assumed column diagonal dominance, the following three
numbers are equal:
Pm• the total encirclements of (−1/ki + j0) made by Gii (jw) for all i, N =
i=1 Ni ,
• the total encirclements of (−1/ki + j0) made by Gershgorin bands,
• the total encirclements of (−1/ki + j0) made by characteristic loci.
Hence, the result follows from the generalized Nyquist Stability Criterion. 
 3.3 Plant Uncertainty and Stability Robustness 99

For G(s) in (3.40), we have Po = 0. It follows from the theorem that

the closed-loop is stable if the net encirclement of (−1/ki + j0) made by the
Gershgorin bands of G is zero. By inspection of Figure 3.13, this is the case
when k1 > 0 and k2 > 0.
Several remarks on Theorem 3.2.8 are made as follows.
• Note that in the theorem, the condition on diagonal dominance has been
imposed. In this case, individual Nyquist plots of gii (s) can be used to
assess closed-loop stability. The design can be carried out loop by loop,
each for one gii (s), like a sequence of SISO systems.
• Due to diagonal dominance, we have allowed K to be diagonal with dif-
ferent diagonal elements so that we can adjust gains for each loop. This is
stronger than that in the multivariable Nyquist theorem where only k = kI
is allowed and the same gain must be used for all loops.
• If (I +GK) is not diagonally dominant, some Gershgorin band may overlap
the origin, the characteristic loci of (I + GK) may or may not encircle the
origin, an hence we cannot infer stability solely by inspecting Gershgorin
bands.

3.3 Plant Uncertainty and Stability Robustness

Most control designs are based on some plant model. A model can never be a
perfect match to a real plant and there must be modelling errors, which are
also called model uncertainties or perturbations. Due to such uncertainties,
a controller which works well for a design model may not necessarily work
equally well for the real plant. It is thus of great importance to represent un-
certainties as accurately as possible, analyze their effects on control systems,
and take them into control design considerations.
To represent uncertainties requires the introduction of singular values.
Singular values are very useful in matrix analysis and control engineering,
and they are a good measure of the “size”of the matrix. Let A be a com-
plex matrix. The positive square roots of the eigenvalues of GH G or GGH ,
whichever has smaller size, are called the singular values of A, are denoted
by σi (A), where AH is the conjugate and transpose of A. A nice thing with
singular values is that we actually need not compute them directly from eigen-
value calculation but through unitary transformation (numerically stable and
efficient), as stated below.
Lemma 3.3.1. (Singular Value Decomposition) Let A ∈ Cp×m .
p×p m×m
There are unitary matrices U ∈ C and V ∈ C such that
A = U ΣV H ,
 
Σ1 0
where Σ = , Σ1 = diag {σ1 , σ2 , · · · , σr }, σ1 ≥ σ2 ≥ . . . ≥ σr ≥ 0, and
0 0
r = min{p, m}.
100 3. Stability and Robustness of Feedback Systems

The following notations are widely adopted:

σ(A) = σmax (A) = σ1 = the largest singular value of A,
and
σ(A) = σmin (A) = σr = the smallest singular value of A.
Let kxk denote the Euclidean vector norm for a vector x. Then the induced
matrix norm from the vector norm, k · k, is defined by
kAxk
kAk = sup ,
x6=0 kxk
and is given by the spectral norm:
kAk = σ(A). (3.41)
Lemma 3.3.2. Suppose that A and B are square matrices. Then
(i) Rank A = number of non-zero singular values of A;
(ii) A is nonsingular if and only if σ(A) > 0;
1
(iii) σ(A−1 ) = σ(A) if A is nonsingular;
(iv) σ(AB) ≤ σ(A)σ(B);
(v) σ(AB) ≥ σ(A)σ(B);
(vi) σ(A + B) ≥ σ(A) − σ(B); and
(vii) σ(AB) < 1 or σ(BA) < 1 if and only if δσ(A) < 1, assuming that
B are all matrices such that σ(B) ≤ δ.
Consider now a transfer function matrix G(s). For a particular s, say,
s = jw for a fixed w, G(jw) is a complex matrix, and the matrix norm,
kG(jw)k = σ(G(jw)), measures the largest gain of the system at this partic-
ular frequency. The largest gain of the system over all frequencies w may be
measured by a operation norm, L∞ − norm:
kGk∞ = sup σ(G(jw)). (3.42)
w

If we plot σ(G(jw)) versus w, then kGk∞ is simply the peak value of

σ(G(jw)).

3.3.1 Unstructured Uncertainty

Let G0 (s) be a nominal transfer matrix, which is a best estimate, in some
sense, of the true plant behavior. Let G(s) denote the true transfer matrix of
the plant. The following are three most commonly used uncertainty models:
• Additive Perturbation: G(s) = G0 (s) + ∆(s), (3.43)
• Input Multiplicative Perturbation: G(s) = G0 (s)(I + ∆(s)), (3.44)
• Output Multiplicative Perturbation: G(s) = (I + ∆(s))G0 (s). (3.45)
 3.3 Plant Uncertainty and Stability Robustness 101

Two kinds of multiplicative perturbations (input or output) are required as

matrix multiplications are not commutative. These uncertainty descriptions
are said to be unstructured if the only restriction on the perturbations ∆ is
on their “size”, which is usually measured by

σ(∆(jw)) ≤ δ(jw), (3.46)

where δ(·) is a positive scalar function. Note that δ represents the absolute
magnitude of model errors in the additive perturbation case and the relative
magnitude in the multiplicative perturbation cases.
The design problem in an uncertain environment consists in selecting a
controller K(s) in a unity feedback system so as to ensure stability as well as
satisfactory performances not only for the nominal plant G0 but also for any
G which belongs to the given uncertain model set. As for the basic stability
requirement, the control system is called robustly stable if it is stable for all
permissible G.

Theorem 3.3.1. (Unstructured Uncertainty) Suppose that K(s) stabi-

lizes the nominal plant G0 and that G and G0 have the same number of
unstable poles. Then, the closed-loop system in Figure 3.10 is robustly stable
for the uncertainty meeting (3.46) if and only if
(i) δ(jw)σ(K(jw)[I + G0 (jw)K(jw)]−1 ) < 1, ∀w ∈ R, for G(s) in (3.43);
(ii) δ(jw)σ(K(jw)G0 (jw)[I + K(jw)G0 (jw)]−1 ) < 1, ∀w ∈ R, for G(s) in
(3.44);
(iii) δ(jw)σ(G0 (jw)K(jw)[I + G0 (jw)K(jw)]−1 ) < 1, ∀w ∈ R, for G(s) in
(3.45).

Proof. We will prove (i) only as other two cases are similar. Since the closed-
loop is stable for ∆(s) = 0 (nominally stabilized), and since G and G0 have
been assumed to have the same number of unstable poles, it follows from The-
orem 3.2.5 that the closed-loop remains stable as long as there is no change
in the number of encirclements of the origin made by det(I + G(jw)K(jw))
with respect to the nominal det(I + G0 (jw)K(jw)), which is equivalent to

det[I + G(jw)K(jw)] 6= 0

for all w ∈ R and permissible G. Bring (3.43) in the above:

det[I + (G0 + ∆)K] = det[I + G0 K + ∆K]
= det[I + ∆K(I + G0 K)−1 ] · det[I + G0 K]
6= 0, (3.47)
which, due to det[I + G0 K] 6= 0, is equivalent to

det[I + ∆K(I + G0 K)−1 ] 6= 0. (3.48)

By Lemma 3.3.2(ii) we can rewrite the above as

102 3. Stability and Robustness of Feedback Systems

σ[I + ∆K(I + G0 K)−1 ] > 0. (3.49)

By Lemma 3.3.2(vi) with A = I, σ(I) = 1 and B = ∆K(I + G0 K)−1 , (3.49)

is implied by

1 − σ(∆K(I + GK)−1 ) > 0,

or σ(∆K(I + G0 K)−1 ) < 1, which, by Lemma 3.3.2(vii), is ensured by Con-

dition (i) in the theorem.
Conversely, suppose that the condition is violated, i.e., for some w∗ , there
holds

δ(jw∗ )σ(K(jw∗ )[I + G0 (jw∗ )K(jw∗ ]−1 ) = 1. (3.50)

Let M = K(I + G0 K)−1 and M (jw) have the singular value decomposition:

M = U ΣV H

with σ1 = σ(M (jw∗ )) at the (1, 1) position of Σ(jw∗ ). Choose ∆ to be

∆ = −δV U H
which is permissible since σ(∆(jw)) ≤ δ(jw). Then, we have
det[I + ∆(jw∗ )M (jw∗ )] = det[I − δ(jw∗ )Σ(jw∗ )]
= det(diag{1 − δ(jw∗ )σ(M (jw∗ )), ?, · · · , ?})
=0
where the last equality is due to (3.50). This in turn, by (3.47), implies

det[I + G(jw∗ )K(jw∗ )] = 0,

for a permissible G = G0 + ∆. Hence jw∗ is a pole of the corresponding

closed-loop system and the system is unstable. 

3.3.2 Structured Uncertainties

In contrast to the unstructured uncertainties given in (3.46), in many cases,

we may trace uncertainties into certain parts or elements of the system and
lead to so-called structured uncertainties. Usually, there will be both struc-
tured and unstructured uncertainties. In order to capture them accurately,
we adopt the following general and standard representation of uncertain sys-
tems. Let the plant have three sets of inputs and outputs.
• 1st set inputs: all manipulated variables.
• 2nd set inputs: all other external signals (disturbances/set points).
• 1st set outputs: all measured variables for feedback.
• 2nd set of outputs: all other outputs whose behaviors are of interest.
 3.3 Plant Uncertainty and Stability Robustness 103

Fig. 3.14. Rearrangement of Uncertain System

The third set of inputs and outputs is novel and comes from uncertainty. For
each uncertainty in the plant (either structured or unstructured), we take it
outside the plant and assign it with one block. Collect them all together as a
special system, which is around the plant and has a block-diagonal structure,
on the diagonal are just those blocks which have been pulled out from inside
the plant. Write

∆(s) = diag {∆1 (s), · · · , ∆n (s)}. (3.51)

where ∆i may be a scalar or a matrix. The resultant representation is shown

in Figure 3.14, noting that the third set of inputs and outputs for plant is
outputs and inputs for uncertainty. By putting suitable weights in loops, we
can always normalize ∆ such that

k∆i k∞ ≤ 1, i = 1, 2, · · · , n. (3.52)

This normalization is used to formulate the standard problem.

If the compensator is already known, we can form a single system Q
representing the closed-loop system consisting of G0 and K with the first set
of inputs and outputs disappeared, as shown in Figure 3.15, for which general
robust stability conditions will be derived below. Define x, v and z according
to Figure 3.15, then one has
      
y v Q11 Q12 v
=Q = .
x z Q21 Q22 z
Suppose that a compensator K(s) stabilizes a nominal model G0 (s) (called
nominal stabilization). Then, Q(s) will be stable.
104 3. Stability and Robustness of Feedback Systems

Fig. 3.15. Standard Representation of Uncertain System

Example 3.3.1. Consider a unity feedback system in Figure 3.10, where G

has element-wise uncertainties:
0 0
|gij (s) − gij (s)| ≤ |aij (s)|.|gij (s)|,
0
with the nominal plant G0 = {gij }. We write
0 0 0
gij (s) = gij + (gij − gij ) := gij + δij ,

to get elementwise the additive uncertainty ∆ = {δij } and its bounds:

0
|δij | ≤ |aij |.|gij | , αij (s).

We can redraw the uncertain system into Figure 3.16, where

e := diag{∆
∆ e2 , ∆
e1 , ∆ e3 , ∆
e4 }
(3.53)
:= diag{δ11 /α11 , δ21 /α21 , δ12 /α12 , δ22 /α22 }
ei k∞ ≤ 1, i = 1, 2, 3, 4.
is the normalized uncertainty with k∆ ♦

Turn back to the standard representation of uncertain system in Figure 3.15.

Suppose nominal stabilization. Our concern is whether or not the system
remains stable for all permissible ∆. Since Q(s) is stable and since Q11 ,
Q12 and Q21 are outside the feedback loop, to know whether the system
will remain stable under permissible perturbations, we only need to check
whether the (positive) feedback combination of Q22 and ∆ will remain stable
for all allowable ∆. In the case of the unstructured uncertainty, or n = 1,
apply Theorem 3.3.1 with assignment of G0 = 0, K = −Q22 , and δ = 1, to
get a necessary and sufficient condition for robust stability:
 3.3 Plant Uncertainty and Stability Robustness 105

Fig. 3.16. Redrawing of Uncertain System into Standard Form

106 3. Stability and Robustness of Feedback Systems

σ(Q22 (jw)) < 1, ∀w ∈ R,

or equivalently

kQ22 k∞ < 1. (3.54)

In general, however, a uncertainty ∆ may have more than one blocks in its
standard representation in Figure 3.15, i.e. n ≥ 2 for ∆ in (3.52). Then the
condition (3.54) is only sufficient for robust stability, because most perturba-
tions which satisfy k∆k ≤ 1 are no longer permissible. (3.54) is usually too
conservative to use for such structured uncertainties.
Let BDδ denote the set of stable, block-diagonal perturbations with a
particular structure, and k∆j k ≤ δ. Suppose nominal stability of the system
in Figure 3.15. It follows from the multivariable Nyquist theorem that the
feedback system can be unstable if and only if

det[I − Q22 (jw)∆(jw)] = 0 (3.55)

for some w and some ∆ ∈ BDδ . We thus define the structured singular value
µ(Q22 (jw)) as follows:
µ(Q22 (jw)) =

0, if det[I − Q22 ∆] 6= 0, for any ∆ ∈ BD∞ ;
(3.56)
1/(min∆∈BD∞ {σ(∆(jw))| det[I − Q22 (jw)∆(jw)] = 0}), otherwise.
In other words, one tries to find the smallest structured ∆ (in the sense of
1
σ(∆)) which makes det(I − Q22 (jw)) = 0, then, µ(Q22 ) = σ(∆) ; If no ∆
meets (3.55), then µ(Q22 ) = 0.
Example 3.3.2. Let
 
2 2
Q22 = .
−1 −1
The smallest (unstructured) full matrix perturbation ∆ which makes det(I −
Q22 ∆) = 0 is
 
0.2 0.2
∆= ,
−0.1 −0.1
and

σ(∆) = 0.316; µ(Q22 ) = 3.162.

On the other hand, the smallest (structured) diagonal matrix perturbation

∆ which makes det(I − Q22 ∆) = 0 is
 
1 10
∆= ,
3 01
and
σ(∆) = 0.333; µ(Q22 ) = 3. ♦
 3.3 Plant Uncertainty and Stability Robustness 107

This example shows that the structured singular value µ(.) depends on
the structure of ∆ and that a larger perturbation in the structured form is
usually allowed than unstructured one, before the matrix becomes singular.
One notes that if ∆ = 0, I − Q22 ∆ = I is nonsingular. Thus, to obtain
µ(Q22 ), one way is to gradually increase σ(∆) until we find I − Q22 ∆ is
singular. Besides, the sequence of Q22 and ∆ in µ-definition does not matter
since

det(I − Q22 ∆) = det(I − ∆Q22 ).

The structured singular value µ(Q) depends on the structure of the set BDδ
as well as on Q. It has the following properties.

Lemma 3.3.3. There hold

(i) µ(αQ) = |α|µ(Q), α scalar;
(ii)µ(I) = 1;
(iii)
µ(Q) = σ(Q), if BDδ has only one block;
(iv)Let diag {Iki } has the same block-diagonal structure as BDδ , define D =
diag {di Iki }, di > 0, then µ(Q) = µ(DQD −1 ) ≤ inf D σ(DQD −1 ); and
(v) If i ≤ 3, i.e., there are no more than three blocks in ∆ and ∆i are
independent of each others, then µ(Q) = inf D σ(DQD −1 ).

With structured singular values, we can define

kQ22 kµ = sup µ(Q22 (jw)),

w

and state a robust stability theorem for structured uncertainty (Doyle,1982).

Theorem 3.3.2. (Structured Uncertainty) The system shown in Figure

3.15 remains stable for all ∆ ∈ BD1 if and only if

kQ22 kµ < 1. (3.57)

One sees that this condition is similar to (3.54) for the unstructured case.
The only difference is that the singular value has been replaced here with
the structured singular value. Under condition (3.57), one can show that
maxω |λmax (Q22 ∆)| < 1 so that (I − Q22 ∆) can not be singular on ω axis and
det(I − Q22 (jw)∆(jw)) 6= 0, and the system is thus stable for all ∆ meeting
k∆j k ≤ 1. If ∆ has one block only, by Lemma 3.3.3(ii), (3.57) reduces to
(3.54). It should be pointed out that the structured singular value is difficult
to compute and we usually use the property (iv) in Lemma 3.6 to overestimate
µ(Q).
Example 3.3.3. Continue Example 3.3.1, we have to find Q22 which meets
x = Q22 z. One sees
108 3. Stability and Robustness of Feedback Systems

 
   
x1 α11 u1 α11 0  
 x2   α21 u1   α21 0 

x= =   =   u1 ,
x3 α12 u2   0 α12  u2
x4 α22 u2 0 α22
  
    
     z1  
 
u1 y1 u1 1010   z2 

= −K = −K G0 + ,
u2 y2 
 u2 0 1 0 1  z3  
 
z4
giving
   
u1 −1 1010
= −(I + KG0 ) K z,
u2 0101
so that
 
α11 0  
 α21 0 
x = −  (I + KG0 )−1 K 1 0 1 0 z.
 0 α12  0101
0 α22
:= Q22 z.
The system remains stable if and only if

kQ22 kµ ≤ 1,

where k · kµ is computed with respect to ∆ = diag{∆1 , ∆2 , ∆3 , ∆4 }, and ∆i

are scalar. ♦

3.3.3 Quantitative Robust Stability Analysis

Theorems 3.3.1 and 3.3.2 are actually special cases of the small gain theo-
rem. Essentially, they all say the same thing: if a feedback loop consists of
stable systems and if the open-loop gain (in a suitable sense) is less than
unity, then the closed-loop system is stable. These results are elegant and
have been widely used in the area of robust control. Unfortunately, it could
be very conservative in many cases. To reveal this conservativeness, one notes
that descriptions in (3.46) for unstructured uncertainty and (3.51) and (3.52)
for structured ones are actually the same for each individual block ∆i which
is restricted by its bound only, and allow much richer uncertainties to occur
than what could exist in reality. In particular, phase changes in the uncer-
tainties are not limited at all, and an infinite phase change is permissible,
which is definitely unrealistic in practice. In real life, both gain and phase
uncertainties would be finite. For example, a common practice in modelling is
that a parametric plant model with some parameter uncertainty is in general
valid over a working frequency range, and may become useless beyond the
 3.3 Plant Uncertainty and Stability Robustness 109

range due to unmoulded dynamics, noise and so on. This implies that phase
uncertainty is much smaller in low frequencies than that in high frequencies
range, and surely finite. For illustration, let us give a concrete example to
show conservativeness of Theorem 3.3.1 and to motivate our sequel develop-
ment.

Example 3.3.4. Let an uncertain plant be described by

1 + δ1
G(s) = , (3.58)
s2 + (1 + δ2 )s + (1 + δ3 )

where δ1 ∈ [−0.5, 0.5], δ2 ∈ [−0.5, 0.5] and δ3 ∈ [−0.5, 0.5]. Suppose that
the controller K(s) = 1. One may easily see that the closed-loop system is
robustly stable, because its characteristic polynomial,

p(s) = s2 + (1 + δ2 )s + 2 + δ1 + δ3 , (3.59)

always has positive coefficients and thus its roots lie in th open left half of
the complex plane.
Now, Let us try the small gain theorem. The nominal plant is
1
G0 (s) = . (3.60)
s2 + s + 1
we may express the uncertainty as an additive one:
1 + δ1 1
∆(s) = − . (3.61)
s2 + (1 + δ2 )s + (1 + δ3 ) s2 + s + 1

Theorem 3.3.1 would say that the feedback system is robustly stable if and
only if

|Q22 (jω)∆(jω)| < 1, ω ∈ [0, ∞) (3.62)

where
−K(s) s2 + s + 1
Q22 (s) := =− 2 . (3.63)
1 + K(s)G0 (s) s +s+2

We have
1 1 + δ1
sup|Q22 (jω)∆(jω)| > lim |Q22 (jω)∆(jω)| = ×| − 1|. (3.64)
ω ω→0 2 1 + δ3
If we choose δ1 = 0.5 and δ3 = −0.5, or the plant is perturbed to
1.5
G(s) = , (3.65)
s2 + s + 0.5
there results in
110 3. Stability and Robustness of Feedback Systems

sup|Q22 (jω)∆(jω)| > 1, (3.66)

ω

violating (3.62) though the system is stable. This shows the conservativeness
of (3.62) that assumes infinite phase change at any frequency, which is not
true in this example. ♦

It should be pointed out that the structured uncertainty and the theory of
structured singular values cater to the block diagonal structure of the uncer-
tainty, and have nothing to do with further information on individual blocks
such as breaking them into magnitude and phase. In the rest of this subsec-
tion, we will acknowledge the importance of limited phase uncertainty and
incorporate such information into the robust stability conditions, in addition
to the existing magnitude information, so as to remove conservativeness and
make the stability criterion exact.
To be specific, we take the SISO unity feedback system for study. If the
plant G(s) and the controller K(s) have no poles in the closed right half
plane, then the Nyquist criterion tells that the closed-loop system is stable
if and only if the Nyquist curve of G(jω)K(jω) does not encircle the critical
point, (−1 + j0). An uncertain plant is always represented by a set of models.
Each model gives rise to a Nyquist curve and the set thus sweeps a band of
Nyquist curves. Then from the Nyquist criterion, the closed-loop system is
robustly stable if and only if the band of the Nyquist cluster for the uncertain
plant and the controller does not encircle the point, (−1 + j0). Conversely,
the system is not robustly stable if and only if there exits a Nyquist curve
among the cluster which encircles −1 point, or there must exist a specific
frequency ω0 at which the frequency response region of the uncertain loop is
such that both

max{|G(jω0 )K(jω0 )|} > 1 (3.67)

G

and

min{arg{G(jω0 )K(jω0 )}} 6 −π (3.68)

G

hold true. Note that

max{|G(jω0 )K(jω0 )|} = |K(jω0 )|max{|G(jω0 )|}, (3.69)

G G

and

min{arg{G(jω0 )K(jω0 )}} = arg{K(jω0 )} + min{arg{G(jω0 )}}.

G G
(3.70)

Thus, we establish the following quantitative robust stability theorem.

 3.3 Plant Uncertainty and Stability Robustness 111

Theorem 3.3.3. Suppose stability of the uncertain plant and the controller.
The uncertain closed-loop system is robustly stable if for any ω ∈ R, either
1
|K(jω)| 6 (3.71)
max{|G(jω)|}
G

or

arg{K(jω)} > −π − min{arg{G(jω)}} (3.72)

G

holds.

For a SISO uncertain system where the controller is unstable, we can

always convert it to the standard representation in Figure 3.15. Suppose
that the nominal plant is stabilized and that the uncertain plant is stable,
too. Then, Q22 (s) and ∆(s) are stable and Theorem 3.3.3 is applicable with
obvious assignment of K(s) = −Q22 (s) and G(s) = ∆(s).

Corollary 3.3.1. For the system in Figure 3.15, if the nominal controlled
system Q22 (s) and the uncertainty ∆(s) are both stable, then the uncertain
closed-loop system remains stable if for any ω ∈ R, either
1
|Q22 (jω)| 6 (3.73)
max{|∆(jω)|}
∆

or

arg{−Q22 (jω)} > −π − min{arg{∆(jω)}} (3.74)

∆

holds.

For analysis of a given uncertain control system, the robust stability cri-
terion in (3.71) and (3.72) can be tested graphically as follows:
(i) Draw Bode plots for magnitude 1/max{|G(jω)|} and phase min{arg
{G(jω)}} respectively, according to the given structure of the uncertain
plant;
(ii) Draw Bode plots for K(jω) in the same diagram; and
(iii) Check if condition (3.71) or (3.72) is satisfied.
Similarly, one can graphically determine the boundaries of the Nyquest band
of a given uncertain plant. Such boundaries will contain the worst case of the
plant which can facilitate robust stability analysis and robust control design.
Obviously, it is more powerful if analytical expressions for the uncertain plant
boundaries can be derived. Though this is difficult in general, it is possible
to do so for a class of typical processes, the uncertain second-order plus dead
time model.
112 3. Stability and Robustness of Feedback Systems

Let us consider
b
G(s) = e−Ls , (3.75)
s2 + a1 s + a2

where the uncertain parameters are b ∈ [b− b+ ], a1 ∈ [a− + − +

1 , a1 ], a2 ∈ [a2 , a2 ]
− + − − − −
and L ∈ [L , L ], b > 0, a1 > 0, a2 > 0 and L > 0. For a fixed ω, the
gain is
p
|G(jω)| = b/ (a1 ω)2 + (a2 − ω 2 )2 . (3.76)

max{|G(jω)|} and min{|G(jω)|} can be obtained as

  q 

 +

 √ − b
, ω ∈ 0, a− 2 ;

 (a1 ω)2 +(a− 2 2
2 −ω ) q
 + q 
b
max{|G(jω)|} = a− ω , ω∈ a2 , a+
−
2 ; (3.77)

 1
 

 q

 +
 √ − 2b + 2 2 , ω ∈ a+
2 ,∞ ;
(a1 ω) +(a2 −ω )

and
  q 

 b− a−
2 +a2
+
√ , ω ∈ 0, 2 ;
(a1 ω) +(a+
+ 2 2 2
2 −ω )
min{|G(jω)|} = q  (3.78)

 b− a−
2 +a2
+
√ , ω∈ 2 ,∞ .
+
(a1 ω) +(ω 2 −a−
2
2 )
2

In a similar way, the phase of the system is given as

arg(G(jω)) = −arg(a2 − ω 2 + ja1 ω) − Lω. (3.79)

It follows that
  q 


−arg(a− 2 + +
2 − ω + ja1 ω) − L ω,ω ∈ 0, a− 2 ;
min{arg(G(jω))} = q 


−arg(ja−
1 ω + a−
2 − ω 2
) − L +
ω, ω ∈ a−
2 , ∞ ;

and
  q 

 + −
−arg(a 2 − ω 2
+ ja1 ω) − L −
ω, ω ∈ 0, a+ 2 ;
max{arg(G(jω))} = q 

 + + 2 − +
−arg(ja 1 ω + a2 − ω ) − L ω, ω ∈ a2 , ∞ .
(3.80)
 3.4 Notes and References 113

0
−1 0 1
10 10 10

−50

−100

−150

−200
−1 0 1
10 10 10

Fig. 3.17. The plot of max{|G(jω)|} and min{arg{G(jω)}} in Example 3.3.5

Example 3.3.5. In Example 3.3.4, the small gain theory fails to decide sta-
bility of the uncertain system given there. We now use Theorem 3.3.3, an
exact robust stability criterion, with the help of the above precise uncertainty
bounds to re-examine the system. Apply (3.77) and (3.80) to get
  √ 

 √ 1.5
, ω ∈ 0, 0.5 ;
 (0.5ω) +(0.5−ω )
2 2 2
√ √ 
1.5
max{|G(jω)|} = 0.5ω , ω∈ 0.5, 1.5 ;

 √ 
√ 1.5
, ω∈ 1.5, ∞ ;
2 2 2
(0.5ω) +(1.5−ω )

and
(  √ 
−arg(0.5 − ω 2 + j1.5ω), ω ∈ 0, 0.5 ;
min{arg(G(jω))} = √ 
−arg(0.5 − ω 2 + j0.5ω), ω ∈ 0.5, ∞ .

with K(jω) = 1, though the gain condition (3.71) is violated when ω < 1.66,
see the gain plot of Figure 3.17. But the phase condition (3.72) will be satisfied
for all frequencies, see the phase plot of Figure 3.17. Thus the uncertain
system is robustly stable. ♦

3.4 Notes and References

The results in this chapter on internal stability using polynomial matrix frac-
tions and frequency response can be found in many linear systems mono-
graphs (Rosenbrock, 1974; Callier and Desoer, 1982; Chen, 1984; Morari and
114 3. Stability and Robustness of Feedback Systems

Zafiriou, 1989), and the exception is Section 1, which is based on Wang et

al. (1999b). The materials on uncertainties and robustness in Section 3 are
standard, and can be found in Maciejowski (1989) and Zhou et al. (1996),
and the exception is the quantitative robust stability criterion, which is due
to Wang et al. (2002).
4. State Space Approach

In this chapter we shall consider a system of the following state space form
ẋ = Ax + Bu,
y = Cx. (4.1)
If x(0) = 0 and the input and the output vectors have the same dimension
m, these are related by the transfer function matrix:
y(s) = G(s)u(s) = C(sI − A)−1 Bu(s), (4.2)
which may be expanded into
y1 (s) = g11 (s)u1 (s) + g12 (s)u2 (s) + · · · + g1m (s)um (s),
y2 (s) = g21 (s)u1 (s) + g22 (s)u2 (s) + · · · + g2m (s)um (s),
.. (4.3)
.
ym (s) = gm1 (s)u1 (s) + gm2 (s)u2 (s) + · · · + gmm (s)um (s).
These equations are said to be coupled , since each individual input influences
all of the outputs. If it is necessary to adjust one of the outputs without
affecting any of the others, determining appropriate inputs u1 , u2 , . . . , um
will be a difficult task in general. Consequently, there is considerable interest
in designing control laws which remove this coupling, so that each input
control only the corresponding output. A system of the form (4.1) is said to
be (dynamically) decoupled (or non-interacting) if its transfer function matrix
G(s) is diagonal and non-singular, that is,
y1 (s) = g11 (s)u1 (s),
y2 (s) = g22 (s)u2 (s),
.. (4.4)
.
ym (s) = gmm (s)um (s).
and none of the hii (s) are identically zero. Such a system may be viewed as
consisting of m independent subsystems. Note that this definition depends
upon the ordering of inputs and outputs, which is of course quite arbitrary.
The definition will be extended to block-decoupling case in Section 4.5.
One example of a system requiring decoupling is a vertical take-off air-
craft. The outputs of interest are pitch angle, horizontal position, and al-
titude, and the control variables consist of three different fan inputs. Since
Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 115-128, 2003.
Springer-Verlag Berlin Heidelberg 2003
116 4. State Space Approach

these are all coupled, the pilot must acquire considerable skill in order to
simultaneously manipulate the three inputs and successfully control the air-
craft. The system may be decoupled using state-variable feedback to provide
the pilot with three independent and highly stable subsystems governing his
pitch angle, horizontal position, and altitude.
It should be pointed out that dynamic decoupling is very demanding. A
signal applied to input ui must control output yi and have no whatsoever
effect on the other outputs. In many cases this requires a complex and highly
sensitive control law, and in other cases it cannot be achieved at all (without
additional compensation). It is useful, therefore, to consider a less stringent
definition which involves only the steady-state behavior of the system.
A system of the form (4.2) is said to be statically decoupled if it is stable
and its static gain matrix G(0) is diagonal and non-singular. This means
that for a step function input u(t) = α1(t), where α = [α1 , α2 , · · · , αm ]T is
constant, the outputs satisfy

limt→∞ y1 (t) = g11 (0)α1 ,

limt→∞ y2 (t) = g22 (0)α2 ,
.. (4.5)
.
limt→∞ ym (t) = gmm (0)αm ,

where gii (0) 6= 0, i = 1, 2, . . . , m. Again, note that the definition depends

upon the ordering of inputs and outputs. A step change in one input ui will
in general cause transients to appear to all of the outputs, but (4.5) ensures
that yj , j 6= i, will be unchanged in a steady state. We will quickly solve
the static decoupling problem in the next section and then concentrate on
dynamic decoupling until section 4.5 from which we discuss block decoupling
problem.

4.1 Static Decoupling

For the system:

ẋ = Ax + Bu, (4.6)
y = Cx + Du, (4.7)
we seek a control policy of the form
u = −Kx + F r, (4.8)
such that the m × m closed-loop transfer function given by
H(s) = (C − DK)(sI − A + BK)−1 BF + DF
= [(C − DK)(sI − A + BK)−1 B + D]F (4.9)
has a diagonal and non-singular H(0), and the control system is stable.
 4.1 Static Decoupling 117

Theorem 4.1.1. The static decoupling problem by state feedback is solvable

if and only if
(i) (A, B)
 is stabilizable;
 and
AB
(ii) rank = n + m.
CD

Proof. If (A, B) is stabilizable, one can find a K such that (A−BK) is stable.
Let us assume that such a K has been found. Consider
    
AB AB In 0
rank = rank
CD CD −K Im
 
A − BK B
= rank
C − DK D
  
In 0 A − BK B
= rank
(C − DK)(−A + BK)−1 Im C − DK D
 
A − BK B
= rank . (4.10)
0 (C − DK)(−A + BK)−1 B + D

Note that (A − BK) is non-singular since it is stable. Thus,

 
AB
rank =n+m
CD
implies that rank[(C − DK)(−A + BK)−1 B + D] = m. With F chosen as
F = [(C − DK)(−A + BK)−1 B + D]−1 , one gets H(0) = I. Hence static de-
coupling is possible. Conversely, if static decoupling is possible, (A, B) should
be stabilizable (trivially). And
rank[H(0)] = rank[(C − DK)(−A + BK)−1 BF + DF ] = m.
This requires that both F and [(C − DK)(−A + BK)−1 B + D] are non-
singular. Thus we have, again by (4.10), that
   
AB A − BK B
rank = rank = n + m,
CD 0 (C − DK)(−A + BK)−1 B + D

from which the necessity follows. 

Design for Static Decoupling. Assume that conditions (i) and (ii) in
Theorem 4.1.1 are satisfied, thus the problem is solvable. We proceed as
follows.
(i) Design a K such that the state feedback system is stable, i.e, det(sI −
A + BK) has all its roots in the left half plane;
(ii) Obtain F as F = [(C − DK)(−A + BK)−1 B + D]−1 ; and
(iii) Form the feedback control law as u = −Kx + F r.
118 4. State Space Approach

Example 4.1.1. Consider the system:

   
0 1 0 1β
ẋ =  0 0 1  x +  0 1  u,
−6 −11 −6 00
 
100
y= .
010
One can easily check that
 
AB
rank = 5,
CD
and the system is stabilizable. Hence static decoupling is possible. Since the
system is stable, let K = 0. Choose
F = [−CA−1 B]−1
 
−6 −17
= (1/36) , when β = 1.
6 11
The resulting closed-loop transfer function is
 
6(s + 6) −s(6s + 25) 1
H(s) = .
6s(s + 6) (11s2 + 66s + 36) (s + 1)(s + 2)(s + 3)6
Clearly, H(0) = I as it should be. ♦

4.2 Dynamic Decoupling

In this section we consider dynamic decoupling of the system:

ẋ = Ax + Bu, x(0) = x0 ,
(4.11)
y = Cx,
where x ∈ Rn , u, y ∈ Rm , A ∈ Rn×n , B ∈ Rn×m , C ∈ Rm×n , by the control
law:
u = −Kx + F r. (4.12)
This problem is called decoupling by state feedback.
When the control law (4.12) is employed, the resultant system becomes
ẋ = (A − BK)x + BF r, (4.13)
y = Cx, (4.14)
and the transfer function matrix is given by
H(s) = C(sI − A + BK)−1 BF. (4.15)
 4.2 Dynamic Decoupling 119

Therefore decoupling by state feedback requires one to find the control ma-
trices K and F which make H(s) diagonal and non-singular. Before treating
the decoupling problem, we first consider the relation between the transfer
function H(s) and that of the plant (4.11).
G(s) = C(sI − A)−1 B. (4.16)

Proposition 4.2.1. The transfer function matrix H(s) in (4.15) is related

to G(s) in (4.16) by
 
H(s) = G(s) I − K(sI − A + BK)−1 B F
 −1
= G(s) I + K(sI − A)−1 B F. (4.17)

Proof. One notes that

H(s) = C(sI − A + BK)−1 BF
= C(sI − A)−1 [(sI − A + BK) − BK](sI − A + BK)−1 BF
= C(sI − A)−1 [I − BK(sI − A + BK)−1 ]BF
= C(sI − A)−1 [B − BK(sI − A + BK)−1 B]F
= C(sI − A)−1 B[I − K(sI − A + BK)−1 B]F
= G(s)[I − K(sI − A + BK)−1 B]F.
Consider now
[I − K(sI − A + BK)−1 B][I + K(sI − A)−1 B]
= I − K(sI − A + BK)−1 (sI − A + BK − BK)(sI − A)−1 B
+K(sI − A)−1 B − K(sI − A + BK)−1 BK(sI − A)−1 B
= I,
which implies that
[I − K(sI − A + BK)−1 B] = [I + K(sI − A)−1 B]−1 , (4.18)
and the proposition is thus proved. 

Proposition 4.2.1 indicates that controlling the system (4.11) with the
control law (4.12) is equivalent to compensating for the system (4.11) serially
by using the compensator
Hc (s) = [I − K(sI − A + BK)−1 B]F, (4.19)
as shown in Figure 4.1. This compensator can be represented by the state
space equations:
ẋc = (A − BK)xc + BF r, (4.20)
yc = −Kxc + F r, (4.21)
where r is the input to the compensator and yc is its output.
120 4. State Space Approach

Fig. 4.1. Series Compensator

Let
 
cT1
 cT2 
 
C= . ,
 .. 
cTm
and define integers σi , i = 1, 2, · · · , m, by
 ÿ

 min j|cTi Aj−1 B 6= 0T , j = 1, 2, · · · , n − 1 ;
σi = (4.22)

n − 1; if cTi Aj−1 B = 0T , j = 1, 2, · · · , n.
To actually evaluate σi , we start with i = 1, cTi = cT1 . Let j = 1, then
cTi Aj−1 B = cT1 A0 B = cT1 B. If cT1 B = 0, then σ1 = 1; otherwise, increase j
to 2, if cT1 Aj−1 B = c1 AB = 0, then σ1 = 2; otherwise, increase j again until
j = n. Repeat this procedure for i = 2, 3, · · · , m.
With σi , i = 1, 2, · · · , m, we can define
 T σ −1 
c1 A 1 B
 cT2 Aσ2 −1 B 
 
B∗ =  .. , (4.23)
 . 
cTm Aσm −1 B
and
 
c1 Aσ1
 c2 Aσ2 
 
C∗ =  .. . (4.24)
 . 
cm Aσm

Theorem 4.2.1. There exists a control law of the form (4.12) to decouple
the system (4.11) if and only if the matrix B ∗ is non-singular. If this is the
case, by choosing
 4.2 Dynamic Decoupling 121

F = (B ∗ )−1 ,
K = (B ∗ )−1 C ∗ ,
the resultant feedback system has the transfer function matrix:
H(s) = diag{s−σ1 , s−σ2 , · · · , s−σm }.
Proof. Necessity is proved first. Noting that
(sI − A)−1 = s−1 I + As−2 + A2 s−3 + · · · , kA/sk < 1,
the i-th row of the transfer function G(s) of (4.16) can be expanded in poly-
nomials of s−1 as
gi (s)T = cTi (sI − A)−1 B = cTi Bs−1 + cTi ABs−2 + · · ·
= cTi Aσi −1 Bs−σi + cTi Aσi Bs−(σi +1) + · · ·
= s−σi (cTi Aσi −1 B + cTi Aσi Bs−1 + cTi Aσi +1 Bs−2 + · · · )
= s−σi {cTi Aσi −1 B + cTi Aσi (sI − A)−1 B}. (4.25)
Using B ∗ of (4.23), G(s) can be written as
 −σ1 
s 0 ··· 0
 . 
 0 s−σ2 . . . ..  ∗
G(s) =   .
 [B + C ∗ (sI − A)−1 B].
 (4.26)
 .. . .
.. .. 0 
0 · · · 0 s−σm
Thus from Proposition 4.2.1 , the transfer function matrix of the feedback
system is given by
H(s) = G(s)[I − K(sI − A + BK)−1 B]F
= diag(s−σ1 , s−σ2 , . . . , s−σm )[B ∗ + C ∗ (sI − A)−1 B]
× [I − K(sI − A + BK)−1 B]F. (4.27)
∗ ∗ −1
For decoupling, H(s) is non-singular, so are [B + C (sI − A) B][I −
K(sI − A + BK)−1 B]F , and its coefficient matrix for s0 , B ∗ F . Thus the
non-singularity of B ∗ is a necessary condition for decoupling.
The sufficiency is proved by constructing the control law to decouple the
system as follows. If B ∗ is non-singular, let K and F be chosen as
K = B ∗−1 C ∗ , (4.28)
∗−1
F =B . (4.29)
It follows from Proposition 4.2.1 that
H(s) = diag (s−σ1 , s−σ2 , . . . , s−σm )[B ∗ + C ∗ (sI − A)−1 B]
×[F −1 + F −1 K(sI − A)−1 B]−1 . (4.30)
Substituting (4.29) and (4.28) into (4.30) gives
H(s) = diag(s−σ1 , s−σ2 , . . . , s−σm ), (4.31)
and the system is decoupled. 
122 4. State Space Approach

The form of the decoupled system (4.31) having s−σi as the diagonal
elements is called an integrator decoupled system.

Example 4.2.1. Consider the system

   
00 11
ẋ = x+ u,
10 00
 
10
y= x.
01
One checks
 
  11  
cT1 A0 B = cT1 B = 10 = 11 =6 0,
00

yielding σ1 = 1; And
 
  11
cT2 A0 B = cT2 B = 0 1 = 0,
00
  
  00 11  
cT2 A1 B = cT2 AB = 0 1 = 11 ,
10 00
giving σ2 = 2. But one has
 T σ −1   T   
c A 1 B c B 11
B ∗ = 1T σ2 −1 = T1 = ,
c2 A B c2 AB 11
which is singular, and the system cannot be decoupled.
It is noted that the B-matrix of the above system is of rank defect. It is
trivial to show from the feedback system transfer function matrix H(s) in
(4.15) that a necessary condition for decoupling is that both C and B have
full rank. ♦

Example 4.2.2. Consider the 2 by 2 system described by

   
−1 0 0 1 0
ẋ =  0 −2 0  x +  2 3  u,
0 0 −3 −3 −3
 
100
y= x.
111
It is readily checked that
 
cT1 A0 B = cT1 B = 1 0 , or σ1 = 1;
   
cT2 B = 0 0 , cT2 AB = 4 3 , or σ2 = 2,

so that
 
10
B∗ = .
43
 4.2 Dynamic Decoupling 123

Since B ∗ is non-singular, the system can be decoupled by state feedback,

u = −Kx + F r. As
 T σ   T   
∗ c1 A 1 c1 A −1 0 0
C = T σ2 = T 2 = ,
c2 A c2 A −1 −2 −3
it follows from Theorem 4.1.1 that
 
1 3 0
F = B ∗−1 = ,
3 −4 1
and
 
−1 0 0
K = B ∗−1 C ∗ = 5 4 .
3 3 3

The resulting closed-loop transfer function is

" #
1 0
H(s) = s . ♦
0 12
s
Example 4.2.3. The linearized equations governing the equatorial motion of
a satellite in a circular orbit is
   
0 1 0 0 00
 3ω 2 0 0 2ω   
ẋ = Ax + Bu =   x +  1 0  u,
 0 0 0 1   0 0
0 −2ω 0 0 01
 
1000
y = Cx = x,
0010
where the outputs and inputs are radial and tangential
   
r ur
y= , u= .
θ uθ
These are coupled by the orbital dynamics, as indicated by the transfer func-
tion matrix
 
1 2ω
2 2 2 2
 s +ω s(s + ω ) 
H(s) = 

.
−2ω s2 − 3ω 2 
s(s2 + ω 2 ) s2 (s2 + ω 2 )
To determine whether this system can be decoupled, we compute
   
cT1 B = 0 0 , cT1 AB = 1 0 , σ1 = 2;
   
cT2 B = 0 0 , cT2 AB = 0 1 , σ2 = 2.
124 4. State Space Approach

Note that
   
cT AB
∗ 10
B = 1T = =I
c2 AB 01
is non-singular. Thus decoupling can be achieved with
 2 
∗−1 ∗ ∗−1 3ω 0 0 2ω
u(t) = −B C x(t) + B v(t) = − x(t) + v(t),
0 −2ω 0 0
which results in the closed-loop system:
   
0 1 0 0 0 0
 0 0 0 0 1 0
ẋ = (A − BB ∗−1 C ∗ )x + Bv =  0
x +   v,
0 0 1 0 0
0 0 0 0 0 1
 
1000
y = Cx = x.
0010
Note that this decoupled system is in the controllable canonical form, and
we may apply additional feedback:
   
k1 k2 0 0 v̄
v = Kx + v̄ = x+ r ,
0 0 k3 k4 v̄θ
to have
   
0 1 0 0 0 0
 k1 k2 0 0 1 0
ẋ = (A − BB C + BK)x + v̄ = 
∗−1 ∗
0
x +   v̄,
0 0 1 0 0
0 0 k3 k4 0 1
 
1000
y = Cx = x.
0010
If we choose the feedback gains k1 = k3 = −1 and k2 = k4 = −2, each of
the two independent subsystems will be stable with a double pole at s = −1.
The transfer function matrix may easily be shown to be
 
1 0
2
 (s + 1) 
H(s) =  .
0 1
2
(s + 1)
This closed-loop system is not only stable (perturbations from the nominal
trajectory will always decay to zero), but adjustments to r and θ can be made
independently. ♦

In general, however, the decoupled system will not be in any canonical

forms. How can pole placement be done together with decoupling?
 4.3 Poles and Zeros of Decoupled Systems 125

4.3 Poles and Zeros of Decoupled Systems

In this section, the control gain matrices K and F are determined so that
the decoupled system has pre-assigned poles. To this end, let
m
X
n̄ = σi
i=1

be the number of the desired feedback system poles to be assigned. We split

them into m subsets, λij , j = 1, 2, · · · , σi for each i, i = 1, 2, · · · , m, and form
σi
X
φi (s) = (s − λij ), i = 1, 2, · · · , m.
j=1

For each i, we express φi (s) as

φi (s) = sσi + γi1 sσi −1 + γi2 sσi −2 + · · · + γiσi ,
and define
φi (A) = Aσi + γi1 Aσi −1 + γi2 Aσi −2 + · · · + γiσi I,
and
   
c∗∗T
1 cT1 φ1 (A)
 c∗∗T   cT2 φ2 (A) 
 2   
C ∗∗ = . = .. .
 ..   . 
c∗∗T
m cTm φm (A)
Note that C ∗∗ = C ∗ if γij = 0 for all i and j, or equivalently, λij = 0, for all
i and j.

Theorem 4.3.1. Suppose that the system (4.11) can be decoupled by state
feedback. If K and F are chosen as
F = (B ∗ )−1 ,
(4.32)
K = (B ∗ )−1 C ∗∗ ,
then the resultant feedback system has the transfer function given by
 
1 1 1
H(s) = diag , ,··· , . (4.33)
φ1 (s) φ2 (s) φm (s)

Proof. Let the i-th row vector of the transfer function matrix of (4.16) be
denoted by giT (s). Multiplying it by (sσi + · · · + γiσi ) yields from (4.25)
126 4. State Space Approach

(sσi + γi1 sσi −1 + · · · + γiσi )giT (s)

= (1 + γi1 s−1 + γi2 s−2 + · · · + γiσi s−σi )
(cTi Aσi −1 B + cTi Aσi Bs−1 + cTi Aσi +1 Bs−2 + · · · )
= cTi Aσi −1 B + (cTi Aσi B + γi1 cTi Aσi −1 B + γi2 cTi Aσi −2 B + · · · )s−1
+(cTi Aσi +1 B + γi1 cTi Aσi B + γi2 cTi Aσi −1 B + · · · )s−2 + · · ·
= cTi Aσi −1 B + c∗∗T
i Bs−1 + c∗∗T
i ABs−2 + · · ·
T σi −1
= ci A B + ci (Is + As + A2 s−3 + · · · )B
∗∗T −1 −2

= cTi Aσi −1 B + c∗∗T

i (sI − A)−1 B. (4.34)
Dividing both sides of (4.34) by (sσi + γi1 sσi −1 + · · · + γiσi ), G(s) is given by
G(s) = diag [(sσ1 + γ11 sσi −1 + · · · + γ1σi )−1 , (sσ2 + γ21 sσ2 −1 + · · · + γ2σ2 )−1 ,
. . . , (sσm + γm1 sσm −1 + · · · + γm σm )−1 ][B ∗ + C ∗∗ (sI − A)−1 B]. (4.35)

Then it follows from (4.17) and (4.35) that

H(s) = diag[(sσ1 + γ11 sσ1 −1 + · · · + γ1σ1 )−1 , . . . , (sσm + γm1 sσm −1 + · · ·
+γmσm )−1 ][B ∗ + C ∗∗ (sI − A)−1 B][F −1 + F −1 K(sI − A)−1 B]−1 . (4.36)
Thus by choosing K and F as (4.32), the transfer function matrix of the
closed loop system is given by (4.33). 

Example 4.3.1. Consider the system:

   
0 0 0 10
ẋ =  0 0 1  x +  0 0  u,
−1 −2 −3 01
 
110
y= x.
001
T
  T
 
We
 have
 c1 B = 1 0 , which yields σ1 =
 1 and c1 A
 = 0 0 1 , and cT2 B =
0 1 , which yields σ2 = 1 and cT2 A = −1 −2 −3 . Therefore, B ∗ and C ∗
are
   
10 0 0 1
B∗ = , C∗ = .
01 −1 −2 −3
Since B ∗ is non-singular, the system can be decoupled. Using Theorem 4.1.1,
the control law has
F = B ∗−1 = I,
 
0 0 1
K = B ∗−1 C ∗ = ,
−1 −2 −3
which will lead to the integrator decoupled closed-loop:
 4.3 Poles and Zeros of Decoupled Systems 127

" #
1 0
H(s) = s .
0 1s
However, if we want to decouple the system to get
H(s) = diag[(s + 1)−1 , (s + 2)−1 ],
then, we have γ11 = 1, γ21 = 2. (4.32) becomes
 T   
c1 A γ11 cT1
K= T +
c2 A γ21 cT2
 T
1c1
= C∗ +
2cT2
     
0 0 1 110 1 1 1
= + = ,
−1 −2 −3 002 −1 −2 −1
and F = I remains the same as before. ♦

In the above example, a 3rd-order system has been decoupled to give a

2nd-order system. Thus a pole-zero cancellation has taken place. In the next
theorem the zero positions are considered in the decoupling.

Theorem 4.3.2. The linear system (4.11) can be decoupled by the feedback
control law (4.12) with the i-th diagonal element of the transfer function of
the decoupled system having a numerator ni (s) if and only if the i-th row of
G(s), denoted by giT (s), is represented by

giT (s) = cTi (sI − A)−1 B = ni (s)c̄Ti (sI − A)−1 B, i = 1, . . . , m, (4.37)

and
 
c̄T1 A(σ̄1 −1) B
 T (σ̄ −1) 
 c̄2 A 2 B 
 
B̄ = 
∗
 ..

 (4.38)
 . 
 
c̄Tm A(σ̄m −1) B

is non-singular, where σ̄i is


 min(j|c̄Ti Aj−1 B 6= 0T , j = 1, 2, · · · , n − 1);
σ̄i =
 n − 1 : c̄T Aj−1 B = 0T , j = 1, 2, · · · , n.
i

Proof. Sufficiency is proved first. Let

 T
c̄1
 .. 
C̄ =  .  .
c̄Tm
128 4. State Space Approach

Then it follows from (4.37), the transfer function matrix G(s) is

G(s) = diag[n1 (s), n2 (s), . . . , nm (s)]C̄(sI − A)−1 B. (4.39)

By Theorem 4.2.1 and (4.38), (A, B, C̄) can be decoupled to give

H(s) = diag [n1 (s), n2 (s), . . . , nm (s)]diag [d−1 −1

1 (s), . . . , dm (s)]. (4.40)

Therefore, the i-th component of the decoupled transfer function is repre-

sented by ni (s)/di (s) and di (s) is determined to have no common factors
with ni (s). Thus sufficiency is proved.
Necessity is proved next. When the diagonal element of the decoupled
system is given by (4.40), the i-th row vector hTi (s) of the decoupled transfer
function is
hTi (s) = cTi (sI − A + BK)−1 BF
= ni (s)/di (s)eTi , (4.41)
which implies that

hTi (s) = ni (s)c̄Ti (sI − A + BK)−1 BF. (4.42)

On the other hand, it follows from (4.17), (4.41) and (4.42) that
cTi (sT − A)−1 B[I + K(sI − A)−1 B]−1 F
= ni (s)c̄Ti (sI − A)−1 B[I + K(sI − A)−1 B]F.
(4.39) is then derived by dividing both sides of the above by [I + K(sI −
A)−1 B]−1 F from the right side. 
The above theorem indicates that the numerator terms in the decoupled
system are given by the properties of the original system and cannot be
changed.

4.4 Notes and References

The problem of decoupling linear time-invariant multivariable systems re-
ceived considerable attention in the system theoretic literature for sev-
eral decades. Much of this attention was directed toward decoupling by
state feedback (Morgan, 1964; Falb and Wolovich, 1967; Hautus and Hey-
mann, 1983; Descusse, 1991; Descusse et al., 1988; Wonham, 1986). More
general block decoupling were investigated by Koussiouris (1979), Koussiouris
(1980), Commault et al. (1991), Descusse (1991), Williams and Antsaklis
(1986). Sections 2 and 3 are based on Falb and Wolovich (1967), Wolovich and
Falb (1969). The algebraic approach has been exclusively taken in this chap-
ter. For the geometric approach, see Wonham (1986), Camarta and Martı́nez-
Garcı́a (2001).
5. Polynomial Matrix Approach

In contrast to the preceding chapter where a state space model of the plant
and state feedback controller are exclusively utilized for decoupling problem,
we will employ unity output feedback compensation to decouple the plant in
this chapter. Polynomial matrix fractions as an input-output representation
of the plant appear a natural and effective approach to dealing with output
feedback decoupling problem. The resulting output feedback controller can
be directly implemented without any need to construct a state estimator.
In particular, we will formulate a general decoupling problem and give some
preliminary results in Section 1. We start our journey with the diagonal de-
coupling problem for square plants in Section 2 and then extend the results
to the general block decoupling case for possible non-square plants in Sec-
tion 3. A unified and independent solution is also presented in Section 4. In
all the cases, stability is included in the discussion. A necessary and suffi-
cient condition for solvability of the given problem is first given, and the set
of all compensators solving the problem is then characterized. Performance
limitations of decoupled systems are compared with a synthesis without de-
coupling. Numerous examples are presented to illustrate the relevant concepts
and results.

5.1 Problem Formulation and Preliminaries

Let R be the field of real numbers, R[s] be the ring of polynomials, R(s) the
field of rational functions, Rp (s) the ring of proper rational functions, and
Rsp (s) the ring of strictly proper rational functions, respectively, all with
coefficients in R. Consider an unity output feedback system shown in Figure
5.1, where G ∈ Rm×l p (s), the set of m × l proper rational matrices, is a given
plant and K ∈ Rl×m p (s) is a compensator. Let D−1 N = G and Nk Dk−1 = K
be, respectively, left and right coprime polynomial matrix fractions. A list of
integers (m1 , m2 , · · · , mv ) is called a partition of m if mi ≥ 1, i = 1, 2, · · · , v,
v
and Σi=1 mi = m. The system in Figure 5.1 is said to be internally stable
(Chapter 3) if det[DDk + N Nk ] has all its roots in C− , the open left half of
the complex plane, and to be block-decoupled with respect to the partition
(m1 , m2 , · · · , mv ) if the closed-loop transfer function matrix between r and
y is nonsingular and block-diagonal, i.e.,
Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 129-185, 2003.
Springer-Verlag Berlin Heidelberg 2003
130 5. Polynomial Matrix Approach

Fig. 5.1. Unity Output Feedback System

T = GK[I + GK]−1
 
T11 0
 T22 
 
= . 
 .. 
0 Tvv
= block diag{T11 , T22 , · · · , Tvv }, (5.1)
with Tii ∈ Rm p
i ×mi
(s) being nonsingular for all i, i = 1, 2, · · · , v. The case
of v = m and m1 = m2 = · · · = mv = 1 is called diagonal decoupling. If
the list of positive integers {mi } is obviously known from the context and no
confusion arises, we denote by Rm×m d [s], Rm×m
d (s), Rm×m
dp (s) and Rm×m dsp (s)
m×m m×m
the set of block diagonal and nonsingular elements in R [s], R (s),
Rm×m
p (s) and Rm×msp (s), respectively. The noninteracting control problem to
be solved in this chapter is stated as follows.

(P). Let a plant G ∈ Rm×l

p (s) and a partition (mi ) be given. Find a con-
troller K ∈ Rl×m
p (s) such that the resulting system in Figure 5.1 is internally
stable and block decoupled.

The plant G has to have rank m in order for T to be nonsingular. There-

fore, throughout the paper, we assume that the plant G has full row rank. We
define zeros and poles of a transfer matrix G in terms of its Smith-McMillan
form (Chapter 2). A pole or zero is called a LHP one if it is in C− , and a RHP
one if it is in C+ , the closed right half of the complex plane. The following
concept of a stability factorization will result in a simplification of (P).

Definition 5.1.1. Let G ∈ Rm×l (s) be of full row rank. A factorization of

G in the form of G = [G+ 0]G− is called a stability factorization if G+ ∈
Rm×m (s) and G− ∈ Rl×l (s) are nonsingular, the zeros and poles of G+ are
precisely the RHP zeros and poles of G, and the zeros and poles of G− are
precisely the LHP zeros and poles of G, counting multiplicities.

Note that G+ and G− may be non-proper, even if G is proper. A stability

factorization can easily be constructed from the Smith-McMillan form of G.
 5.1 Problem Formulation and Preliminaries 131

Lemma 5.1.1. Any full row rank matrix G ∈ Rm×l (s) has a stability fac-
torization G = [G+ 0]G− . The matrix G+ is unique up to a multiplication
by an unimodular polynomial matrix from the right. In the case of square
G, G− is unique up to a multiplication by an unimodular polynomial matrix
from the left.

Proof. It follows from Chapter 2 that there are unimodular polynomial

matrices U1 and U2 such that

G = U1 SU2 ,

where S = [diag{ri } 0] with ri ∈ R(s) is the Smith-McMillan form of G.

Factorize each ri as ri+ and ri− such that all zeros and poles of ri+ are in
C+ , while all zeros and poles of ri− lie in C− . We then obtain a stability
factorization of G as G = [G+ 0]G− , where

G+ = U1 diag{ri+ },

and

G− = block diag{diag{ri− }, I}U2 .

Consider now any two stability factorizations

G = [G+ − + −
1 0]G1 = [G2 0]G2 .

We have

[G+ + − − −1
1 0] = [G2 0]G2 (G1 ) .

Define U = G− − −1
2 (G1 ) and partition it as
 
U11 U12
U=
U21 U22

with U11 ∈ Rm×m (s), so that

[G+ + +
1 0] = [G2 U11 G2 U12 ]

and

G+ +
1 = G2 U11 . (5.2)

Note that all the distinct poles of U11 must be the poles of U = G− − −1
2 (G1 )
which is stable. So U11 can only have stable poles if any, and by (5.2) these
poles must be also the poles of G+1 because they cannot be cancelled by G2 .
+
+
But G1 cannot have any stable poles by definition, and thus U11 must not
have any pole. By a similar argument using G+ + −1 −1
2 = G1 U11 , U11 must not
132 5. Polynomial Matrix Approach

have any pole either. Hence, U11 is unimodular. For square G, [G+ 0] reduces
to G+ only and U11 = U . U G− −
1 = G2 and the proof is completed. 
Note that for nonsquare full row rank G, G− is highly non-unique: its last
l − m rows are arbitrary subject to the constraint that no poles and zeros
are introduced. The following special stability factorization will be required
later.
Corollary 5.1.1. There exist unique square polynomial matrices Ng+ and
Dg+ and a matrix G− such that G+ + + −1
g := Ng (Dg ) and G− define a stability
factorization, Ng is in the Hermite column form and Dg+ is in the Hermite
+

row form.
Proof. Let G = [G+ − +
1 0]G1 be a stability factorization and G1 = N (D )
+ + −1

be a coprime polynomial matrix fraction. There exists a unique unimodular

polynomial matrix U such that Ng+ := N + U is in the required form (Theorem
2.2.2 of Chapter 2). Moreover, there exists a unique unimodular polynomial
matrix V such that Dg+ := V (D+ U ) is in the required form. 
Let us introduce further notations. Consider a rational matrix R and as-
sume that Dl−1 Nl and Nr Dr−1 are coprime polynomial matrix fractions such
that Dl−1 Nl = Nr Dr−1 = R. The McMillan degree δ(R) of R (Definition
2.3.2 of Chapter 2) is given by δ(R) := deg(det(Dl )) = deg(det(Dr )). The
McMillan RHP degree of R is defined to be the number of all the RHP poles
of R (counting multiplicities), and is given by δ + (R) := deg(det(Dl+ )) =
deg(det(Dr+ )), where Dl = Dl+ Dl− and Dr = Dr+ Dr− are stability factoriza-
tions. It is obvious that δ + (G) = δ + (G+ ) = δ(G+ ) for a stability factorization
G = [G+ 0]G− . Let G and K be two rational matrices of compatible size. The
series connection GK is said to have no pole-zero cancellations if the equality
δ(G)+δ(K) = δ(GK) holds. Similarly, the equality δ + (G)+δ + (K) = δ + (GK)
means that GK has no RHP pole-zero cancellations.
Consider (P) now. Assume first that (P) is solvable, then the closed-loop
transfer matrix given in (5.1) is nonsingular and block-diagonal. Its inverse
is

T −1 = I + (GK)−1 .

Thus, we have GK ∈ Rm×m d (s). The internal stability of the closed-loop

system ensures that GK has no RHP pole-zero cancellations. Conversely,
if there is a K ∈ Rl×m (s) such that GK ∈ Rm×m d (s) and it has no RHP
pole-zero cancellations, then we can always choose a set of stable poly-
nomial matrices Pi ∈ Rmi ×mi [s] such that GK block diag{Pi−1 } is strictly
proper and K block diag{Pi−1 } is proper. It follows from Chapter 3 that
there is a proper rational matrix block diag{Ri } such that it internally sta-
bilizes GK block diag{Pi−1 }. From the above construction, it is obvious that
K̄ = K block diag{Pi−1 Ri } is a solution to (P). Therefore, we have estab-
lished the following theorem.
 5.2 Diagonally Decoupling for Square Plants 133

Theorem 5.1.1. For a plant G ∈ Rm×l p (s), (P) is solvable if and only if
there is a K ∈ Rl×m (s) such that GK is block-diagonal and nonsingular, and
has no RHP pole-zero cancellations.

Theorem 5.1.1 means that (P) is equivalent to the decoupling problem

for a plant G by precompensator K without RHP pole-zero cancellations in
GK, a great simplification from the closed-loop problem to the open-loop
problem. It is noted that in the latter case, K is not required to be proper.
The skew prime polynomial matrices were introduced by Wolovich (1978)
and have appeared in linear system theory for solving various output regu-
lation and/or tracking problems. It turns out that our solvability conditions
for (P) also include a skew primeness one.

Definition 5.1.2. Let a full row rank N ∈ Rm×l [s] and a nonsingular D ∈
Rm×m [s] be given. If there are N̄ ∈ Rm×l [s] and D̄ ∈ Rl×l [s] such that

DN = N̄ D̄,

where D and N̄ are left coprime, and N and D̄ right coprime, then N and D
are called as externally skew prime and N̄ and D̄ as internally skew prime.

Algorithms for constructing N̄ and D̄ from N and D have been given in

Wolovich (1978) and Wang, Sun and Zhou (1989). In particular, if N and D
have no common zeros, then they are externally skew prime. In the sequel
development, we will make frequent use of the following lemma.

Lemma 5.1.2. Let A, B, C, D be all polynomial matrices such that AB −1 C =

D. If B −1 C (resp. AB −1 ) is coprime, then AB −1 (resp. B −1 C) is a polyno-
mial matrix.

Proof: If AB −1 is not a polynomial matrix, let ÃB̃ −1 be a coprime poly-

nomial matrix fraction of AB −1 with deg(detB̃) > 0, then D = AB −1 C =
ÃB̃ −1 C with ÃB̃ −1 and B̃ − C both coprime. It follows (Theorem 2.13 of
Chapter 2) that D is a rational matrix with its Mcmillan degree δ(D) =
deg(detB̃) nonzero, contradicting the assumption that D is a polynomial
matrix. The proof is completed. 

5.2 Diagonally Decoupling for Square Plants

Let us consider first the simplest case of (P): diagonal decoupling for square
plants, which is abbreviated as (PS) for ease of frequent future reference.
Then G should be square and nonsingular. Our task, by Theorem 5.1.1, is to
find a nonsingular K such that GK is nonsingular and diagonal and has no
RHP pole-zero cancellations.
134 5. Polynomial Matrix Approach

5.2.1 Solution

The solution K is shown in the following theorem to be representable as the

product of the plant inverse and a diagonal nonsingular matrix.
Theorem 5.2.1. (PS) is solvable if and only if there is a T ∈ Rm×m
d (s)
such that GK with K = G−1 T has no RHP pole-zero cancellations.
Proof: If (PS) is solvable, it follows from Theorem 5.1.1 that there is a
K such that GK has no RHP pole-zero cancellations and GK = T for some
T ∈ Rm×m
d . Then K = G−1 T . On the other hand, if there is a T ∈ Rm×md such
that GK with K = G−1 T has no RHP pole-zero cancellations, then GK = T
is diagonal and nonsingular, and by Theorem 5.1.1, (PS) is solvable. The
proof is completed. 
The concept of strict adjoints of a polynomial matrix was introduced by
Hammer and Khargonekar (1984). Let P, Pra ∈ Rm×m [s] be nonsingular, Pra
is said to be a right strict adjoint of P whenever the following conditions are
satisfied: (i) P Pra ∈ Rm×m
d [s], and (ii) if any nonsingular Q ∈ Rm×m [s] is
m×m
such that P Q ∈ Rd [s], then Pra is a left divisor of Q, i.e., Q = Pra Q1 for
some Q1 ∈ Rm×m [s].
Lemma 5.2.1. Let P ∈ Rm×m [s] be nonsingular, then P has a right strict
adjoint.
A right strict adjoint of P can be constructed as follows. The inverse
P −1 of P is first computed. Let (P −1 )i be the i-th column of P −1 and ci be
the least common multiple of denominators of all elements in (P −1 )i , then
it can be shown (see the general case in Lemma 5.3.1 and its proof) that
Pra = P −1 diag{ci } is a right strict adjoint of P . The concept of left strict
adjoints can be defined in a dual way and the dual result to Lemma 5.2.1
also holds.
Consider now the plant G in details. By Lemma 5.1.1, there is a stability
factorization for the nonsingular G:
G = G+ G− . (5.3)
We further factorize G+ as
G+ = N + (D+ )−1 , (5.4)
where N + and D+ are nosingular polynomial matrices and right coprime. By
Lemma 5.2.1, N + has a right strict adjoint Nra
+
such that N + Nra
+
∈ Rm×m
d [s].
+ +
Further, if Nra and D are externally skew prime, there exists a pair of
+
internally skew prime polynomial matrices N̄ra and D̄+ such that
D+ Nra
+ + +
= N̄ra D̄ . (5.5)
+
Let D̄la be a left strict adjoint of D̄+ , we can now state the main solvability
conditions for (PS).
 5.2 Diagonally Decoupling for Square Plants 135

+
Theorem 5.2.2. (PS) is solvable if and only if Nra and D+ are externally
+ + + +
skew prime and N Nra and D̄la D̄ are coprime.

Proof. Sufficiency: If the conditions in the theorem are satisfied, then we

take K as

K = G−1 T, (5.6)

where T ∈ Rm×m
d (s) is given by the coprime polynomial matrix fraction
+ + −1
T = N + Nra
+
(D̄la D̄ ) . (5.7)

This K can be reduced as

+ + −1
K = (G− )−1 D+ (N + )−1 N + Nra
+
(D̄la D̄ )
+ + −1
= (G− )−1 N̄ra D̄+ (D̄la D̄ )
− −1 + + −1
= (G ) N̄ra (D̄la ) . (5.8)
Then from (5.3) and (5.4) we obtain

δ + (G) = deg(det(D+ )). (5.9)

(5.7) and (5.8) imply that

+
δ + (T ) = deg(det(D̄+ )) + deg(det(D̄la )), (5.10)

and
+
δ + (K) ≤ deg(det(D̄la )). (5.11)

It follows from the skew primeness equation (5.5) that

deg(det(D̄+ )) = deg(det(D+ )). (5.12)

Collecting (5.9)-5.12) yields

δ + (T ) ≥ deg + (G) + deg + (K), (5.13)

which means that GK has no RHP pole-zero cancellations since GK = T .

By Theorem 5.2.1, (PS) is solvable.
Necessity: If (PS) is solvable, then by Theorem 5.2.1, there is a T ∈
Rm×m
d (s) such that GK with K = G−1 T has no RHP pole-zero cancellations.
Let T = NT DT−1 , with NT , DT ∈ Rm×m
d [s], be coprime, K then becomes

K = G−1 T = (G− )−1 D+ (N + )−1 NT DT−1 . (5.14)

We claim that no RHP pole-zero cancellations in GK imply that (N + )−1 NT

is a polynomial matrix. Suppose, conversely, that (N + )−1 NT is a rational
matrix. Let N1+ be a greatest common left divisor of N + and NT , then
136 5. Polynomial Matrix Approach

(N2+ )−1 NT 2 with N2+ = (N1+ )−1 N + and NT 2 = (N1+ )−1 NT is a coprime
polynomial matrix fraction and (N2+ )−1 is not cancelled in G−1 T . Therefore,
there must be such a representation

(Ñ2+ )−1 Q = (G− )−1 D+ (N2+ )−1

with det(Ñ2+ ) = det(N2+ ) that Ñ2+ is a divisor of the denominator of a left

coprime polynomial matrix fraction of K, that is,

K = (DK Ñ2+ )−1 NK

with DK Ñ2+ and NK left coprime. But then this unstable Ñ2+ will disappear
in GK because
−1
GK = [(Ñ2+ )−1 Q(N1+ )−1 ]−1 [(DK Ñ2+ )−1 NK ] = N1+ Q−1 DK NK .

This contradicts the assumption of no RHP pole-zero cancellations in GK.

Thus, we have

(N + )−1 NT = P

or equivalently

NT = N + P

for some polynomial matrix P . Since NT ∈ Rm×m

d [s], we obtain

NT = N + Nra
+
P1 , (5.15)

where Nra+
is a right strict adjoint of N + and P1 ∈ Rm×m
d [s]. Substituting
(5.15) into (5.14) gives
+ −1
K = (G− )−1 D+ Nra D T P1 .
+
The coprimeness of NT and DT implies that Nra and DT are right coprime.
+ −1
Construct a dual left coprime polynomial matrix fraction from Nra DT as

D̃T−1 Ñra
+ + −1
= Nra DT (5.16)

with
+ +
det(Ñra ) = a1 det(Nra ), a1 ∈ R. (5.17)

K becomes

K = (G− )−1 D+ D̃T−1 Ñra

+
P1 .

Similarly, no RHP pole-zero cancellations in GK ensure that there is a P̃2 ∈

Rm×m [s] such that
 5.2 Diagonally Decoupling for Square Plants 137

D̃T = P̃2 D+ . (5.18)

Substituting (5.18) into (5.16) yields

+ −1 + +
DT = P̄2 (N̄ra ) D Nra , (5.19)

where
+ −1 + −1
P̄2 (N̄ra ) = (Ñra ) P̃2 (5.20)

is right coprime with

+ +
det(N̄ra ) = a2 det(Ñra ), a2 ∈ R, (5.21)

and
+
N̄ra and D+ are left coprime. (5.22)

+
Equations (5.21) and (5.22) hold due to the left coprimeness of D̃T and Ñra .
+
Because DT is a polynomial matrix and P̄2 and N̄ra are right coprime, it
follows from (5.19) that
+ −1 + +
(N̄ra ) D Nra = D̄+ (5.23)

for some polynomial matrix D̄+ . Equation (5.23), together with (5.17), (5.21),
+
and (5.22), implies that Nra and D+ are externally skew prime and N̄ra +
and
+
D̄ just constitute their dual internally skew prime pair. With (5.23), (5.19)
is simplified into

DT = P̄2 D̄+ .

Again, it follows from DT ∈ Rm×m

d [s] that
+ +
DT = P2 D̄la D̄ , (5.24)
+
where D̄la is a left strict adjoint of D̄+ and P2 ∈ Rm×m
d [s]. From (5.15) and
+ +
(5.24), one sees that N + Nra +
and D̄la D̄ are, respectively, divisors of coprime
NT and DT and thus they are certainly coprime. The proof is completed. 
Specializing Theorem 5.2.2 to the case of no coinciding RHP pole and
zeros, we obtain the following corollary.

Corollary 5.2.1. For a nonsingular plant G, if det(N + ) and det(D+ ) are

coprime, that is, the plant has no RHP poles coinciding with zeros, then (PS)
is solvable.

Proof. From the inverse matrix construction, we obtain N + adj{N + } =

det(N + )I. It follows from the definition of strict adjoints that
138 5. Polynomial Matrix Approach

N + Nra
+
P = det(N + )I (5.25)

for some P ∈ Rm×m

d [s]. Hence, when det(N + ) and det(D+ ) are coprime, so
+
are det(Nra ) and det(D+ ) and in this case Nra
+
and D+ are externally skew
prime. From the skew primeness equation (5.5), one has

det(D̄+ ) = a1 det(D+ ), a1 ∈ R. (5.26)

A similar equation to (5.25) also holds for D̄+ as

+ +
P 0 D̄la D̄ = det(D̄+ )I (5.27)

for some P 0 ∈ Rm×m

d [s]. Under (5.25), (5.26) and (5.27), the coprimeness of
+ +
det(N + ) and det(D+ ) implies that N + Nra
+
and D̄la D̄ are also coprime and,
by Theorem 5.2.2, (PS) is solvable. 
Apart from the solvability conditions, one needs to characterize decou-
pling compensators and achievable loop maps. The following corollary is pre-
sented for this purpose.
Corollary 5.2.2. If a nonsingular plant G satisfies the conditions of Theo-
rem 5.2.2, then
(i) all the compensators which decouple G with internal stabilizability are
given by
+ −1
+
K = (G− )−1 N̄ra (D̄la ) K2 , (5.28)

where K2 ∈ Rm×m d (s) is such that K1 K2 with K1 = N + Nra

+ + + −1
(D̄la D̄ ) has
no RHP pole-zero cancellations.
(ii) all the achievable open-loop maps GK under decoupling with internal
stabilizability can be expressed by

GK = K1 K2 ,

where K1 and K2 are the same as in (i).

Proof. It follows from the proof of Theorem 5.2.2 that every K which decou-
ples G without RHP pole-zero cancellations has the form

K = G−1 NT DT−1 ,
+ +
where NT = N + Nra +
P1 , DT = P2 D̄la D̄ for some P1 , P2 ∈ Rm×m
d [s], and
NT and DT are coprime. Define K2 = P1 P2−1 , then NT DT−1 = K1 K2 . The
coprimeness of NT and DT implies that K1 K2 has no RHP pole-zero cancel-
lations. Simple calculations yield
+ −1
K = G−1 K1 K2 = (G− )−1 N̄ra
+
(D̄la ) K2 ,

which is (5.28). In addition, we have GK = K1 K2 as well, and hence the

corollary follows. 
 5.2 Diagonally Decoupling for Square Plants 139

From the point of view of closed-loop system design, all remains to do is to

construct a diagonal compensator K̃ for the decoupled plant G̃ = GG−1 K1 =
K1 . It becomes a sequence of SISO design problems for which effective meth-
ods are available.
The decoupling compensators constructed so far usually result in LHP
pole-zero cancellations (G− (G− )−1 ) so that arbitrary pole assignment is not
achievable. However, by imposing the stronger constraint, no pole-zero cancel-
lations in GK, instead of only RHP ones, we can directly obtain the following
result and the proof is omitted because of its similarity to that of Theorem
5.2.2.
Theorem 5.2.3. For a nonsingular plant G, the following hold true.
(i) There is a proper compensator in a unity output feedback system such
that it decouples G and achieves arbitrary closed-loop pole assignment if and
only if Nra and D are externally skew prime and N Nra and D̄la D̄ are co-
prime, where N D−1 is a right coprime polynomial matrix fraction of G, N̄ra
and D̄ are a dual internally skew prime pair of the externally skew prime Nra
and D such that DNra = N̄ra D̄, and Nra and D̄la are, respectively, right and
left strict adjoints of N and D̄.
(ii) Under the conditions of (i), all the compensators which decouple G
and preserve arbitrary pole assignability are given by

K = N̄ra (D̄la )−1 K2 ,

where K2 ∈ Rm×m d (s) is such that K1 K2 with K1 = N Nra (D̄la D̄)−1 has no
pole-zero cancellations.
(iii) All the achievable open-loop maps under decoupling with arbitrary
pole assignability can be expressed by

GK = K1 K2 ,

where K1 and K2 are defined as in (ii).

We now present two examples to illustrate our results.
Example 5.2.1. Consider the plant:
 1 1 
s+1 s+2
G= 1 s ,
(s−1)(s+1) (s−1)(s+2)

which is the transfer matrix of the state-space model example in Linnemann

and Maier (1990) for α = 1. G is factorized as

G = G+ G− ,

where
 
1 1
G+ = 1 s ,
s−1 s−1
140 5. Polynomial Matrix Approach

and
 1 
0
G− = s+1
1 .
0 s+2

From G+ , we construct one of its right coprime polynomial matrix fractions

as
  −1
s−1 0 s−1 −1
G+ = N + (D+ )−1 = .
1 1 0 1

One sees that G has a common zero and pole at s = 1, which is unstable.
Compute (N + )−1 as
 1 
+ −1 s−1 0
(N ) = −1 ,
s−1 1

then,
 
+ + −1 1 0
Nra = (N ) diag{(s − 1), 1} = .
−1 1
+
Obviously, Nra and D+ are externally skew prime since det(Nra
+
) ∈ R has no
root at all. One easily checks that
 
s −1
D̄+ =
−1 1
+
and N̄ra = I2 satisfy the skew primeness equation

D+ Nra
+ + +
= N̄ra D̄ ,

and they constitute a dual skew prime pair. A left strict adjoint of D̄+ is
similarly obtained as
 
+ 1 1
D̄la = .
1 s

Hence, we have
 
s−1 0
N + Nra
+
= ,
0 1

and
 
+ + s−1 0
D̄la D̄ = .
0 s−1

They are not coprime. (PS) for this plant is thus not solvable. ♦
 5.2 Diagonally Decoupling for Square Plants 141

Example 5.2.2. Consider now a 3 × 3 plant:

 s 1

s+1 s+2 0
 
G= 1
 s(s+1)
1
s(s+2) 0 .

1 1 1
(s−1)(s+3) (s−1)(s+4) (s−1)(s+5)

A stability factorization of G is given by

  1 
s 1 0 s+1 0 0
1  
+ −  1
0   1
0 
G=G G =s s  0 s+2 .
1 1 1 1
0 0 s−1 s+3 s+4 s+5

Again, G has a zero coinciding with a pole at s = 1. We could go step

by step following Theorem 5.2.2. But, one notes that this G+ has a special
block-diagonal structure and it enables us to decompose the original (PS)
into several smaller ones. In fact, the proof of Theorem 5.2.1 shows that
G− of a plant plays no role in solvability of (PS), and we can always take
K as K = (G− )−1 K̃ to cancel it without affecting the solvability of (PS).
Furthermore, if G+ has a block-diagonal structure G+ = block diag{G+ i } and
there is a solution K̃i to (PS) for each G+
i , then K = (G − −1
) block diag{ K̃i }
will be a solution to (PS) for the original plant G. Therefore, we obtain the
following sufficient condition for (PS). ♦
Theorem 5.2.4. If a nonsingular plant G has a stability factorization G =
G+ G− with G+ = block diag{G+ +
i } and for each Gi (PS) is solvable, then
(PS) for G is also solvable. In particular, if G = diag{gi+ }, then (PS) for
+

G is solvable.
Now turn back to the example, we have

G+ = block diag{G+ +
1 , G2 },

where G+
2 = 1/(s − 1) is scalar, for which (PS) is of course solvable, and
 
s 1
G+1 = 1 1 .
s s

The zero of G+
1 at s = 1 is different from its pole at s = 0, and by Corollary
5.2.1, (PS) for G+1 is also solvable. It follows from Theorem 5.2.4 that (PS)
for G is thus solvable. ♦

5.2.2 An Alternative Condition

This subsection presents an alternative solvability condition which involves

transfer function matrices only and requires no coprime factorizations of poly-
nomial matrices.
142 5. Polynomial Matrix Approach

Given an m×m nonsingular proper rational transfer function matrix G, let

the set of all the common RHP poles and zeros of G be Γ = {γ1 , γ2 , . . . , γn }.
Let δp (G) denote the McMillan degree of a transfer function G at the RHP
pole p, i.e., multiplicity of the pole at p. For simplification, we denote δ p (G) =
δp (G) + δp (G−1 ). Gd is called a decoupler for G if GGd is decoupled.

Definition 5.2.1. Gd,min is called a minimal C + - decoupler for G if (i)

GGd,min is decoupled; (ii) For any decoupler Gd for G, there holds δ γ (Gd,min )
≤ δ γ (Gd ) for any γ ∈ Γ .

Lemma 5.2.2. (Existence and uniqueness of Gd,min ) For any non-

singular G , (i) there exists at least a Gd,min ; (ii) the set of the Gd,min which
have different zero and pole structure at γ ∈ Γ is finite, each number is unique
up to right multiplication by a decoupled matrix which has neither pole nor
zero at any γ ∈ Γ , and all the members have the same δ γ (Gd,min ).

Proof. (i) Obviously, Gd = G−1 is a decoupler for any m × m nonsingular

G. One can see that Gd D is still a decoupler for any decoupled D ∈ Rm×m d .
For each γ ∈ Γ , we perform the following search to find Gd,min from an
initial decoupler Gd . Let lp (resp. lz ) be the greatest multiplicity of the
pole of all the elements of Gd (resp. G−1 d ) at γ ∈ Γ . Suppose that lj , j =
1, 2, · · · , m, are integers, D = diag{(s−γ)l1 , (s−γ)l2 , · · · , (s−γ)lm } and L =
{(l1 , l2 , · · · , lm )| − lz ≤ li ≤ lp , i = 1, 2, · · · , m}. If (l1 , l2 , · · · , lm ) ∈
/ L, then
there exists at least an lj∗ such that either (a) lj∗ > lp or (b) lj∗ < −lz holds.
Suppose the case (a) (the case (b) can be treated with obvious changes). Then,
Gdiag{(s − γ)l1 , · · · , (s − γ)lj−1 , (s − γ)lp , (s − γ)lj+1 , · · · ,∗(s − γ)lm } cannot
have any pole-zero cancellation with diag{1, · · · , 1, (s − γ)lj −lp , 1, · · · , 1} and
thus δ r (GD) with (l1 , l2 , · · · , lm ) = (l1 , · · · , lj−1 , lj∗ , lj+1 , · · · , lm ) is strictly
greater than δ r (GD) with (l1 , l2 , · · · , lm ) = (l1 , · · · , lj−1 , lp , lj+1 , · · · , lm ). It
thus follows that a minimum δ r (GD) will only come from some D with
(l1 , l2 , · · · , lm ) ∈ L if it exists. Further, the combinations of such lj in the
range L are finite and the unique minimum δ γ (Gd D) must be achieved with
one or a few combinations of lj in the range and the corresponding D will
yield Gd,min = Gd D.
(ii) From the above construction, one sees that the number of such Gd,min
for a given initial Gd is finite and all of them have the same δ̄r (Gd,min ). Let
now G̃d,min be an arbitrary minimal C + -decoupler for G, which may not
come from the above search. Then, GG̃d,min = D1 , and any Gid,min from
the above search also meets GGid,min = D2 , where D1 and D2 are decoupled.
This gives rise to G̃d,min = Gid,min D3 for a decoupled D3 = D1 D2−1 . Factorize
D3 as D3 = D3γ D3γ̄ , where decoupled D3γ contains all the poles and zeros
of D3 at all γ ∈ Γ and D3γ̄ has no pole or zero at any γ ∈ Γ . Then, we
have G̃d,min = (Gid,min D3γ )D3γ̄ and δ̄γ (G̃d,min ) = δ̄γ (Gid,min D3,γ ). It follows
that Gid,min D3γ is a minimal C + -decoupler. Further, Gid,min D3γ falls into the
 5.2 Diagonally Decoupling for Square Plants 143

search in (i) and thus it is itself a member of Gd,min , say, Gjd,min . Therefore,
we conclude that G̃d,min = Gjd,min D3γ . The proof is completed. 

Remark 5.2.1. Note that the locations of poles of a multivariable system must
be the same as those of its elements (Chapter 2), and that the zeros of G
are the poles of G−1 . Thus, we can find the common RHP zeros and poles
of G by examining the poles of the elements of G and G−1 . This observation
together with the proof of Lemma 5.2.2 naturally gives rise to the following
constructive procedure for finding Gd,min from G.

Step 1: Calculate Gd = G−1 and determine the set Γ from pole locations
of elements of G and G−1 . If Γ = φ, i.e. empty set, take Gd,min =
Gd ; otherwise, proceed to step 2.
Step 2: For each γ ∈ Γ , find lpγ and lzγ from pole multiplicity of elements
of Gd and G−1 l1 l2
d . Let Dγ = diag{(s−γ) , (s−γ) , · · · , (s−γ) },
lm

tabulate δ γ (Gd Dγ ) for all possible combinations of lj such that

−lzγ ≤ lj ≤ lpγ and determine the minimum δ γ (GDγ ) and the
corresponding Dγ∗ .
Step 3: Set Gd,min = Gd Dγ∗1 Dγ∗2 · · · Dγ∗n .

Example 5.2.3. Let

" 1 1 #
s+1 s+2
G= 1 s
,
(s−1)(s+1) (s−1)(s+2)

which is the same as in Example 5.2.1. It follows that

" s(s+1) #
s−1 −(s + 1)
G−1 = ,
s+2
− s−1 (s + 2)

for which we can easily find that the only common RHP pole and zero of G
is at 1 i.e., Γ = {1}. Take Gd = G−1 and obviously we have lpγ = 1 and
lzγ = 1. For D = diag{(s − 1)l1 , (s − 1)l2 }, where −1 ≤ l1 , l2 ≤ 1, the search
for Gd,min is tabulated in Table 5.1, from which two Gd,min are found as
" s(s+1) s+1
#
s−1 − s−1
Gd,min1 = ,
s+2 s+2
− s−1 s−1

and
 
s(s + 1) −(s + 1)
Gd,min2 = .
−(s + 2) (s + 2)

If we start our procedure with a very different G̃d :

144 5. Polynomial Matrix Approach

" s+1 #
1 − s+2
G̃d = s+2
,
− s(s+1) 1

with lpγ = 0. And we have

" s s(s+1) #
s−1 (s+2)(s−1)
G̃−1
d = s+2 s
,
(s−1)(s+1) s−1

so that lzγ = 1. For D = diag{(s − 1)l1 , (s − 1)l2 } with −1 ≤ l1 , l2 ≤ 0, the

search for Gd,min is tabulated in Table 5.2, from which two Gd,min are found
as
" s+1 #
1 − s+2
G̃d,min1 = s+2
, (5.29)
− s(s+1) 1

and
" 1 s+1 #
s−1 − (s+2)(s−1)
G̃d,min2 = s+2 1
. (5.30)
− s(s+1)(s−1) s−1

Tables 5.1 and 5.2 verify that δ γ (Gd,min ) is unique for a given G. Although
two sets of the Gd,min obtained above look different, actually they are related
to each other by
 −1
s(s + 1) 0
Gd,min1 = G̃d,min2 ,
0 (s + 2)

and
" 1
#−1
s(s+1) 0
Gd,min2 = G̃d,min1 1
,
0 s+2

which is in agreement with (ii) of Lemma 5.2.2. ♦

Lemma 5.2.3. If Gd is a solution to (PS) for G, then there exists a Gd,min

such that (i) Gd = Gd,min D, where D ∈ Rm×m d is decoupled; and (ii)
δγ (Gd ) = δγ (Gd,min ) + δγ (D), where γ ∈ Γ .
+ +
Proof. Take Gd,min = (G− )−1 N ra (Dla )−1 as in (5.28) and the lemma follows
from Corollary 5.2.2. 

Theorem 5.2.5. For a nonsingular G, (PS) is solvable if and only if there

exists a Gd,min which has no pole-zero cancellation with G at any γ ∈ Γ ⊂
C+ .
 5.2 Diagonally Decoupling for Square Plants 145

Table 5.1. Search for Gd,min in Example 5.2.3

 
(s − 1)l1
(l1 , l2 ) K = Gd δ 1 (K)
(s − 1)l2

" s(s+1) s+1

#
(s−1)2
− s−1
(−1, −1) s+2 s+2 3
− (s+1) 2 s−1

" s(s+1) #
(s−1)2
−(s + 1)
(−1, 0) s+2 3
− (s−1) 2 s+2

" s(s+1) #
(s−1)2
−(s + 1)(s − 1)
(−1, 1) s+2 4
− (s−1) 2 (s − 1)(s + 2)

" #
s(s+1) s+1
s−1 − s−1
(0, −1) s+2 s+2 1
− s−1 s−1

" #
s(s+1)
s−1 −(s + 1)
(0, 0) s+2 2
− s−1 s+2

" #
s(s+1)
s−1 −(s + 1)(s − 1)
(0, 1) s+2 3
− s−1 (s − 1)(s + 2)

" #
s+1
s(s + 1) − s−1
(1, −1) s+2 2
−(s + 2) s−1

 
s(s + 1) −(s + 1)
(1, 0) 1
−(s + 2) s + 2

 
s(s + 1) −(s + 1)(s − 1)
(1, 1) 2
−(s + 2) (s + 2)(s − 1)

Proof. Sufficiency: By the definition of Gd,min , GGd,min is decoupled. The

theorem will follow if Gd,min has no pole-zero cancellation with G at any
ρ ∈ C+ but ρ ∈ / Γ . Then, such ρ can be a pole or a zero of G but not both. If
ρ be a pole (resp. zero) of G, then ρ must be a zero (resp. pole) of Gd,min for
1 l
the cancellation to occur. Form Gd = (s−ρ) l Gd,min (resp. (s−ρ) Gd,min ) for a

sufficiently large integer l such that Gd no longer has any zero (resp. pole) at
ρ. This Gd has no pole-zero cancellation with G at ρ. Repeat this elimination
of pole-zero cancellation for all possible ρ and the resultant Gd will have no
pole-zero cancellation with G at ρ ∈ C+ but ρ ∈ / Γ.
Necessity: It follows from Theorem 5.1.1 that if (PS) is solvable, then
there is a Gd such that GGd is decoupled and meets

δγ (G) + δγ (Gd ) = δγ (GGd ), γ ∈ Γ. (5.31)

146 5. Polynomial Matrix Approach

Table 5.2. Search for Gd,min in Example 5.2.3

 
(s − 1)l1
(l1 , l2 ) K = Gd δ 1 (K)
(s − 1)l2

" s+1
#
1
s−1 − (s+2)(s−1)
(−1, −1) s+2 1 1
− s(s+1)(s−1) s−1

" #
1 s+1
s−1 − s+2
(−1, 0) s+2 2
− s(s+1)(s−1) 1

" s+1
#
1 − (s−1)(s+2)
(0, −1) s+2 1 2
− s(s+1) s−1

" #
1 − s+1
s+2
(0, 0) s+2 1
− s(s+1) 1

By Lemma 5.2.3, for Gd there exists a Gd,min such that Gd = Gd,min D and
δγ (Gd ) = δγ (Gd,min ) + δγ (D). (5.32)
Substituting (5.32) into (5.31) yields
δγ (G) + δγ (Gd,min ) + δγ (D) = δγ (GGd,min D), (5.33)
implying that G and Gd,min have no pole-zero cancellation at γ ∈ Γ . This
completes the proof. 
Example 5.2.4. Consider the same G as in Example 5.2.3 with Γ = {1}. For
Gd,min1 in (5.29), one sees
" 1 1 #" # " s−1 #
s+1 s+2 1 − s+1
s+2 s(s+1) 0
GGd,min1 = 1 s s+2
= 1
,
(s−1)(s+1) (s−1)(s+2) − s(s+1) 1 0 s+2

δ1 (G) = 1, δ1 (Gd,min1 ) = 0, and δ1 (GGd,min1 ) = 0, δ1 (G) + δ1 (Gd,min1 ) 6=

δ1 (GGd,min1 ). There is a pole-zero cancelation between G and Gd,min1 at
γ = 1.
For Gd,min2 in (5.30), it follows that
" 1 1 #" 1 s+1 #
s+1 s+2 s−1 − (s+2)(s−1)
GGd,min2 = 1 s s+2 1
(s−1)(s+1) (s−1)(s+2) − s(s+1)(s−1) s−1
" 1 #
s(s+1) 0
= 1
,
0 (s+2)(s−1)
δ1 (G) = 1, δ1 (Gd,min2 ) = 1, and δ1 (GGd,min1 ) = 1, so that δ1 (G) +
δ1 (Gd,min1 ) 6= δ1 (GGd,min1 ). There is a pole-zero cancellation between G
 5.2 Diagonally Decoupling for Square Plants 147

and Gd,min2 at γ = 1 too. Thus, none of Gd,min can meet the condition in
Theorem 5.2.5 and so (PS) for this plant is not solvable. ♦
Example 5.2.5. Let
" s 1 #
s+1 s+2
G(s) = 1 1
.
s(s+1) s(s+2)

Choose
 
s+1
s−1 − s(s+1)
s−1
Gd = G−1 =  .
s+2 s2 (s+2)
− s−1 s−1

It is readily seen that the only common RHP pole and zero of G(s) is 1,
lzγ = 0 and lpγ = 1. From Table 5.3, one can see that Gd,min is not unique.
For
 
s+1 s(s+1)
s−1 − s−1
Gd,min1 =  2
,
s+2 s (s+2)
− s−1 s−1
we have
" # 
 
s 1 s+1
− s(s+1)
GGd,min1 =
s+1 s+2
 s−1 s−1
= 10 ,
1 1 s+2
− s−1 s2 (s+2) 01
s(s+1) s(s+2) s−1

δ1 (G) = 0, δ1 (Gd,min1 ) = 1, and δ1 (GGd,min1 ) = 0, so δ1 (G) + δ1 (Gd,min1 ) 6=

δ1 (GGd,min1 ). There is a pole-zero cancelation between G and Gd,min1 at
γ = 1.
For
" #
s + 1 −s(s − 1)
Gd,min2 = ,
−(s + 2) s2 (s + 2)
we have
 s 1    
s+1 s+2 s + 1 −s(s + 1) s−1 0
GGd,min2 = 1 1 = ,
s(s+1) s(s+2) −(s + 2) s2 (s + 2) 0 s−1

δ1 (G) = 0, δ1 (Gd,min2 ) = 0, and δ1 (GGd,min2 ) = 0, so that δ1 (G) +

δ1 (Gd,min2 ) = δ1 (GGd,min2 ). There is no pole-zero cancelation between G
and Gd,min2 at γ = 1, and (PS) for this G(s) is solvable. ♦

Example 5.2.6. Let

 s 1 s

s−1 − s−1 s−1
G(s) =  −1 1 −s  .
0 0 s−1
148 5. Polynomial Matrix Approach

Table 5.3. Search for Gd,min in Example 5.2.5

 
(s − 1)l1
(l1 , l2 ) K = Gd δ 1 (K)
(s − 1)l2

" s+1 s(s+1)

#
s−1 s−1
(0, 0) s+2 s2 (s+2)
1
− s−1 s−1

" s(s+1)
#
s+1 − s−1
(1, 0) s2 (s+2)
2
−(s + 2) s−1

" #
s+1
s−1 −s(s − 1)
(0, 1) s+2 2
− s−1 s2 (s + 2)

 
s + 1 −s(s + 1)
(1, 1) 1
−(s + 2) s2 (s + 2)

From Table 5.4, there are two Gd,min . For

 1 
1 s 0
 
Gd,min1 =  1 1 s ,
001
we calculate
 s 1 s

  
s−1 − s−1 s−1 1 1s 0 1 0 0
 
−s   1 1 0  =  0 s 0 ,
s−1
GGd,min1 =  −1 1
0 0 s−1 0 0 1 0 0 s − 1

δ1 (G) = 1, δ1 (Gd,min1 ) = 0, and δ1 (GGd,min1 ) = 0, δ1 (G) + δ1 (Gd,min1 ) 6=

δ1 (GGd,min1 ). There is a pole-zero cancelation between G and Gd,min1 at
γ = 1.
For
 1 1

s−1 s(s−1) 0
 
Gd,min2 =  1 1
s , s−1 s−1
0 0 1
we have
 1   1 
  1
0
1
− s−1 ss s−1 s(s−1) 0 0
s−1
s−1  1   s−1 
GGd,min2 
= −1 1   1
s
−s  s−1  =  0 s 0 ,
1
s−1
0 0 s−1 0 0 s−1
0 0 1
δ1 (G) = 1, δ1 (Gd,min2 ) = 1, and δ1 (GGd,min2 ) = 1, δ1 (G) + δ1 (Gd,min2 ) 6=
δ1 (GGd,min2 ). There is a pole-zero cancelation between G and Gd,min2 at
γ = 1, too. So (PS) for the plant is not solvable. ♦
 5.3 Block Decoupling for General Plants 149

Table 5.4. Search for Gd,min in Example 5.2.6

 
(s − 1)l1
(l1 , l2 , l3 ) K = Gd  (s − 1)l2  δ 1 (K)
(s − 1)l3
 
1 1
s−1 s(s−1)
0
 1 1 s 
(−1, −1, −1)  s−1 s−1 s−1  2
1
0 0 s−1

 1 1 
s−1 s(s−1)
0
(−1, −1, 0)  1 1
s 1
s−1 s−1
0 0 1
 1 1 
− s−1 s 0
(−1, 0, −1)  1
1 s 3
s−1
0 0 1

 1 1 
s−1 s 0
(−1, 0, 0)  1
1 s 2
s−1
0 0 1
 
1
1 s(s−1)
0
 1 s 
(0, −1, −1) 1 s−1 s−1  3
1
0 0 s−1

 1 
1 s(s−1)
0
(0, −1, 0) 1 1
s 2
s−1
0 0 1
 
1 1s 0
1 1 s 
(0, 0, −1) s−1 2
1
0 0 s−1

 
1 1s 0
(0, 0, 0) 1 1 s 1
0 0 1

5.3 Block Decoupling for General Plants

In the preceding section, we have established necessary and sufficient con-
ditions for the solvability of the diagonal decoupling problem with internal
stability for unity output feedback systems with square plants. The condi-
tions are only concerned with G+ of a given plant, and they consist of the
+
externally skew primeness of Nra and D+ and the coprimeness of N + Nra +
+ +
and D̄la D̄ over the ring of polynomials. The problem, however, is solved
only for square plants. A difficulty arising from a nonsquare plant is that its
right inverses are not unique and Theorem 5.2.1 is no longer valid in this
case. Another generalization needed is to consider block-diagonalization with
internal stability. These problems form the original (P) and will be solved in
the present section.
150 5. Polynomial Matrix Approach

Proposition 5.3.1. Let G ∈ Rm×l sp (s) be a given plant of full row rank and
G = [G+ 0]G− be a stability factorization. Then (P) is solvable if and only if
there exists a K1 ∈ Rm×m (s) such that G+ K1 is block diagonal, nonsingular
and has no pole-zero cancellations, i.e.,
G+ K1 ∈ Rm×m
d (s), (5.34)
and
δ(G+ ) + δ(K1 ) = δ(G+ K1 ). (5.35)
Proof. Assume that K solves (P). Then by Theorem 5.1.1 GK is block diag-
onal and has no RHP pole-zero cancellations. Let Im be the m × m identity
matrix and
K1 := [Im 0]G− K.
Then GK = G+ K1 (which implies (5.34)) and δ + (K) ≥ δ + (K1 ). Moreover,
G+ K1 has no RHP pole-zero cancellations, because
δ + (G+ ) + δ + (K1 ) ≤ δ + (G) + δ + (K) = δ + (GK) = δ + (G+ K1 ).
Since G+ has only RHP zeros and poles and G+ and K1 are both square,
this means that G+ K1 has no zero-pole cancellations at all, which is (5.35).
Conversely, if K1 satisfies (5.34) and (5.35), then
 
K1
K := (G− )−1 P −1 C (5.36)
0
solves (P), where P ∈ Rm×m
d [s] is stable and such that (G− )−1 [K1T 0]T P −1
m×m
is proper, and C ∈ Rdp (s) internally stabilizes G+ K1 P −1 . 
Proposition 5.3.1 simplifies (P) into the following open-loop block de-
coupling problem with stabilizability for a square and nonsingular rational
matrix having only RHP zeros and poles. Note again that K1 is not required
to be proper.

(PO). Let a nonsingular G+ ∈ Rm×m (s) and a partition be given. Find

a K1 ∈ Rm×m (s) such that G+ K1 is block diagonal, nonsingular and has no
pole-zero cancellations, i.e., (5.34) and (5.35) are satisfied.

Example 5.3.1. Consider the plant given by

 s −1 s 1

s−1 s−1 (s−1)(s+1) s+1
 
G=
 −1 1 −s
s+1 0 

s−1
0 0 s+1 0
 
  1 1 0 −1 −1
1 0 0 0 
1s s −1 
= 0 s − 1 0 0 
0 0 s + 1 0  .
0 0 s−1 0
0 0 0 s+1
 5.3 Block Decoupling for General Plants 151

Its pole and zero polynomial is given by (s − 1)(s + 1)2 and (s − 1)2 , respec-
tively. A stability factorization is defined by
 1

 s  1 0 0 s+1
−1 s  
s−1 s−1 s−1  0 1 0 1 
+   −  s+1 
G =  −1 1 −s  , G =  .
 0 0 1
0 
0 0 s−1  s+1 
1
0 0 0 s+1

By Proposition 5.3.1, (P) for G is equivalent to (PO) for G+ for every

partition. ♦

5.3.1 Solvability

The notion of strict adjoints plays a vital rule in solving the diagonal decou-
pling problem. Its generalization to the block case is obviously required to
solve the general problem (PO).

Definition 5.3.1. Let P ∈ Rm×m [s] be a nonsingular polynomial matrix. A

polynomial matrix Prs ∈ Rm×m [s] is said to be a right strict block adjoint of
P whenever the following conditions are satisfied: (i) P Prs ∈ Rm×m
d [s]; and
(ii) any Pr ∈ Rm×m [s] satisfying P Pr ∈ Rm×m
d [s] is a right multiple of Prs ,
i.e., Pr = Prs Pd for some polynomial matrix Pd ∈ Rm×md [s].

A right strict block adjoint of P can be determined by partitioning P −1

into

P −1 = [P̃1 P̃2 · · · P̃v ],

where P̃i ∈ Rm×mi (s). Let Ai ∈ Rm×mi [s] and Bi ∈ Rmi ×mi [s] be right
coprime such that P̃i = Ai Bi−1 , i = 1, 2, · · · , v. The following lemma shows
that

A := [A1 A2 · · · Av ] (5.37)

is a right strict block adjoint of P .

Lemma 5.3.1. Any nonsingular polynomial matrix P has a right strict block
adjoint with respect to every partition. A particular right strict block adjoint
of P is given by Prs = A, as defined in (5.37). For a fixed partition, right
strict block adjoints of P are uniquely determined up to multiplication by a
block diagonal unimodular matrix from the right.

Proof. By the above construction there holds P A = B with

B := block diag{B1 , B2 , · · · , Bv } ∈ Rm×m

d [s].
152 5. Polynomial Matrix Approach

Assume now that a nonsingular Pr ∈ Rm×m [s] satisfies P Pr = Q =

block diag{Qi } ∈ Rm×m
d [s]. Then the equality:

Pr = P −1 Q = [A1 B1−1 Q1 A2 B2−1 Q2 · · · Av Bv−1 Qv ],

implies that Ai Bi−1 Qi is a polynomial matrix for each i. Since Ai Bi−1 is

coprime, by Lemma 5.1.2, Pi := Bi−1 Qi is a polynomial matrix. This means
that Pr = APd with Pd = block diag{Pi } ∈ Rm×m d [s]. Therefore, A is a right
strict block adjoint of P .
0
Let Prs and Prs be both right strict block adjoints of P . Then there exist
m×m 0 0
Pd1 , Pd2 ∈ Rd [s] such that Prs = Prs Pd1 and Prs = Prs Pd2 . This implies
Prs = Prs Pd2 Pd1 , so that Pd1 and Pd2 are both unimodular. 
We will also need left strict block adjoints which are defined, like Defini-
tion 5.3.1 with the obvious modifications. A dual result to Lemma 5.3.1 will
then also hold.

Example 5.3.2. Consider the following nonsingular polynomial matrix P with

its inverse,
 s 
  −1
0
1 1 0 s−1
 −1
s−1

P =  1 s 0  , P −1 =  1
 s−1 s−1 0  .

0 s 1 s −s
s−1 s−1 1

Let the partition be defined by m1 = 1 and m2 = 2. Then

 s 
 
s−1
 −1  s
P̃1 =     −1
 s−1  = −1 (s − 1) = A1 B1 ,
−1

s s
s−1

and
 
−1
0  

s−1
 −1 0  −1
s−1 0
P̃2 = 

1
s−1

0 =  1 0  = A2 B2−1 .
0 1
−s −s 1
s−1 1

Thus a right strict block adjoint Prs (1, 2) with respect to the partition (1, 2)
is given by
 
s −1 0
Prs (1, 2) =  −1 1 0  .
s −s 1

Using the same approach, a right strict block adjoint Prs (2, 1) with respect
to the partition (2, 1) can be computed to be
 5.3 Block Decoupling for General Plants 153

 
1 −1 0
Prs (2, 1) =  0 1 0  .
0 −s 1

We now turn back to (PO) as defined before and choose right coprime
polynomial matrices N + and D+ such that

G+ = N + (D+ )−1 . (5.38)

The matrix N + is nonsingular, and, by Lemma 5.3.1, it has a right strict block
+
adjoint Nrs . For the special case that D+ is the identity matrix (which means
+
that G is stable), K1 = Nrs solves the problem. Using the same approach
for the general case, we have to guarantee that the order of multiplication of
+
Nrs and (D+ )−1 can be appropriately interchanged. Let us assume that Nrs +

and D+ are externally skew prime. Then there exist N̄rs +

and D̄+ such that
+
Nrs (D̄+ )−1 = (D+ )−1 N̄rs
+
, (5.39)
+
with both fractions coprime. Let D̄ls be a left strict block adjoint of D̄+ and
+ + −1
K1 := N̄rs (D̄ls ) . Then we have
+ + −1
G+ K1 = N + Nrs
+
(D̄ls D̄ ) ,

which is block diagonal. Let us assume that the above fraction is coprime.
Then there holds
+
δ(G+ K1 ) = deg(det(D̄ls )) + deg(det(D̄+ ))
+
= deg(det(D̄ls )) + deg(det(D+ ))
≥ δ(K1 ) + δ(G+ ).
Hence there are no pole-zero cancellations, and K1 solves (PO). The follow-
ing proposition shows that the assumptions made for the above solution of
the problem are also necessary for solvability of (PO).

Proposition 5.3.2. Let a nonsingular G+ ∈ Rm×m (s) and a partition be

+
given. (PO) is solvable if and only if Nrs and D+ are externally skew prime
+ +
and, moreover, N + Nrs
+
and D̄ls D̄ are right coprime.

Proof. Sufficiency has been shown above. The proof of necessity is more in-
volved, but follows the same lines as the proof of Theorem 5.2.2. At those
places where that proof uses the equality Nd Dd−1 = Dd−1 Nd for diagonal poly-
nomial matrices we have to introduce block diagonal and coprime polynomial
matrices with Nd Dd−1 = D̃d−1 Ñd . The details are omitted. 
Propositions 5.3.1 and 5.3.2 directly imply the following solvability con-
dition for the block-decoupling problem with stability using unity output
feedback.
154 5. Polynomial Matrix Approach

Theorem 5.3.1. Let G ∈ Rm×l sp (s) be a given plant of full row rank and
G = [G+ 0]G− be a stability factorization. Then (P) is solvable if and only
+ + +
if Nrs and D+ are externally skew prime and, moreover, N + Nrs
+
and D̄ls D̄
are right coprime.
Let us recapitulate the main steps to check solvability of (P) and to
compute a controller K solving (P): obtain a stability factorization G =
[G+ 0]G− using the Smith-McMillan form of the plant; determine a polyno-
mial fraction as in (5.38), using the Smith-McMillan form of G+ , for instance;
+
construct a right strict block adjoint Nrs of N + as described in Lemma 5.3.1;
check if Nrs and D are externally skew prime and determine D̄+ and N̄rs
+ + +
+ +
satisfying (5.39); determine a left strict block adjoint D̄ls of D̄ as described
+ +
in Lemma 5.3.1; check if N + Nrs +
and D̄ls D̄ are right coprime; determine
K1 = N̄rs (D̄ls ) ; find a stable polynomial matrix P ∈ Rm×m
+ + −1
d [s] such that
(G− )−1 [K1T 0]T P −1 is proper; determine C ∈ Rm×m dp (s) which internally
+ −1
stabilizes the block diagonal G K1 P ; calculate K from (5.36).
Example 5.3.3. Consider the plant of Example 5.3.1. (P) is solvable for G if
and only if (PO) is solvable for
  −1
1 0 0 1 1 0
G+ =  0 (s − 1) 0  1 s s  = N + (D+ )−1 .
0 0 (s − 1) 0 0 1
Note that G+ has a zero of multiplicity 2 at s = 1 coinciding with a single pole
at s = 1. Since N + is diagonal, we can take Nrs +
= I for any partition. Hence
+ + +
Nrs and D are externally skew prime and N̄rs = I and D̄+ = D+ satisfy
(5.39). In order to determine the solvability of (PO) for this example, we
+ + + +
need only to check if N + Nrs +
= N + and D̄ls D̄ = Dls D are right coprime,
+
where Dls is a left strict block adjoint of D . We have (D+ )T = P , where
+
+ + + T T
P is given in Example 5.3.2, and Dls D = ((D+ )T (Dls ) ) . Hence, for the
partitions (2, 1) and (1, 2), there holds
 
1 1 0
+
Dls (2, 1)D+ = (Prs (2, 1))T D+ =  0 s − 1 0  ,
0 0 1
and
 
s−1 0 0
+
Dls (1, 2)D+ = (Prs (1, 2))T D+ =  0 s − 1 0 .
0 0 1
+
One easily sees that N + and Dls (2, 1)D+ are right coprime but N + and
+ +
Dls (1, 2)D are not. Hence, by Theorem 5.3.1, (P) is solvable for the par-
tition (2, 1) but not solvable for the partition (1, 2). For the partition (2, 1),
+
the controller K1 = (Dls (2, 1))−1 solves (PO), and the controller K defined
by (5.36) solves (P). ♦
 5.3 Block Decoupling for General Plants 155

5.3.2 Controller Parameterization

The following proposition characterizes the set of all compensators solving

(PO). Its proof is analogous to that of Corollary 5.2.2 and is therefore omit-
ted.

Proposition 5.3.3. Let G+ ∈ Rm×m (s) be nonsingular and let (PO) be

solvable. Then K1 solves (PO) if and only if
+ + −1
K1 = N̄rs (D̄ls ) Kd ,

where Kd ∈ Rm×md (s) has no pole-zero cancellations with Qd = N + Nrs +

×
+ + −1 +
(D̄ls D̄ ) . An open-loop transfer function matrix Q = G K1 under block
decoupling with internal stabilizability is achievable if and only if Q = Qd Kd .

The main difficulty in generalizing this result to a parameterization of

controllers solving (P) comes from the fact that a non-square controller K
will allow considerable additional degrees of freedom as compared with con-
trollers solving (PO). To illustrate this effect, consider the following simple
example.
1 1
Example 5.3.4. Consider G = [ s−1 0] with G+ = s−1 and G− = I2 . G+ is
already decoupled, so that a controller K1 solves (PO) if and only if it has no
4s
RHP pole-zero cancellations with G+ . For example, K̂1 = s−1 solves (PO).
T
Now, all controllers K = [K̂1 K2 ] with stable K2 clearly solve (P). However,
4s 1 T
there are also unstable K2 such that K solves (P), e.g., K = [ s−1 s−1 ] .
The following theorem shows that unstable poles of K2 do necessarily come
from K1 . Note that some important closed-loop transfer matrices, such as
K(I + GK)−1 , depend on K2 . Therefore, a parameterization of K2 is of
relevance, although it does not influence GK. ♦

Theorem 5.3.2. Let G = [G+ 0]G− be a stability factorization of a full row

rank plant G ∈ Rm×l l×m
sp (s) and let (P) be solvable. Then K ∈ Rp (s) solves
(P) if and only if
 
− −1 K1
K = (G ) ,
K2
+ −1
where K1 has the form K1 = N̄rs +
(D̄ls ) Kd for some Kd ∈ Rm×m
d (s) stabi-
lizing Qd = N Nrs (D̄ls D̄ ) , and K2 has the form K2 = K̃2 D1−1 for some
+ + + + −1

stable K̃2 ∈ R(l−m)×m (s) and a polynomial matrix D1 such that N1 D1−1 is a
right coprime polynomial matrix fraction of K1 .

Proof. Suppose that a K satisfies the conditions of the theorem. Then GK =

G+ K1 = Qd Kd ∈ Rm×m d (s) and the closed-loop system is input-output stable
as Kd stabilizes Qd . This will imply internal stability if GK has no unstable
156 5. Polynomial Matrix Approach

pole-zero cancellations. Since there are no unstable pole-zero cancellations in

Qd Kd and K1 Kd−1 solves (PO) for G+ , we have
δ + (GK) = δ + (Qd Kd )
= δ + (Qd ) + δ + (Kd )
= δ + (G+ (K1 Kd−1 )) + δ + (Kd )
= δ + (G+ ) + δ + (K1 Kd−1 ) + δ + (Kd )
≥ δ + (G) + δ + (K1 ).
Moreover, due to the stability of (G− )−1 and K̃2 ,
 
− −1 N1
+ +
δ (K) = δ ((G ) D1−1 ) ≤ deg + (D1 ) = δ + (K1 ).
K̃2
Hence δ + (GK) ≥ δ + (G) + δ + (K), which means that there are no unstable
pole-zero cancellations in GK.
Conversely, suppose that K solves (P) for G, and define K1 := [I 0]G− K
and K2 := [0 I]G− K. Then GK = G+ K1 and
δ + (GK) = δ + (G) + δ + (K). (5.40)
The proof of Proposition 5.3.1 shows that K1 solves (PO) for G+ , so
+ + −1
that, by Proposition 5.3.3, K1 has the form K1 = N̄rs (D̄ls ) Kd where
m×m
Kd ∈ R d (s) has no pole-zero cancellations with Qd . This implies that
Kd stabilizes Qd . Hence it remains to verify the properties of K2 . The equa-
tion (5.40) and δ + (G+ K1 ) = δ + (G+ ) + δ + (K1 ) imply that δ + (K1 ) = δ + (K),
which reduces to
 
K1
δ + (K1 ) = δ + ( ).
K2
Let Ñ and D̃ be right coprime such that
   
K1 Ñ1
= Ñ D̃−1 = D̃−1 .
K2 Ñ2
Since the fraction on the right hand side of Ñ1 D̃−1 = N1 D1−1 is coprime, it
follows that D̃ = D1 D̃2 for some polynomial matrix D̃2 . The matrix D̃2 is
stable because
 
+ + + + K1
deg (det(D̃2 )) = deg (det(D̃)) − deg (det(D1 )) = δ ( ) − δ + (K1 ) = 0,
K2

where deg + denotes the degree of the unstable part. Hence, K2 = Ñ2 D̃−1 =
K̃2 D1−1 with K̃2 = Ñ2 D̃2−1 being stable. 
It is well known that the achievable performance of a system is essentially
determined by the RHP open-loop poles and zeros. Therefore, in view of our
study in the following section, it is of interest to determine the RHP poles
and zeros which are necessarily introduced by decoupling, i.e., the RHP poles
+ + −1
and zeros of the coprime fraction K̃1 = N̄rs (D̄ls ) .
 5.3 Block Decoupling for General Plants 157

+ +
Corollary 5.3.1. The polynomials det(N̄rs ) and det(D̄ls ) divide
+ m−1 + m−1
[det(N (s))] and [det(D (s))] , respectively. Moreover, the controller
+ + −1
K̃1 = N̄rs (D̄ls ) is stable (respectively, minimum phase) if and only if Dg+
(respectively, Ng+ ) as defined in Corollary 5.1.1 is block diagonal.
+ + +
Proof. We have det(N̄rs ) = det(Nrs ), where Nrs is a right strict block ad-
+ +
joint of N . From the equation N (adj(N )) = (det(N + )I it follows that
+

adj(N + ) is a multiple of Nrs+

and that det(adj(N + )) = (det(N + )m−1 . Thus
det(N̄rs ) divides [det(N (s))]m−1 . Moreover, Ng+ is block diagonal if and
+ +
+
only if Nrs is unimodular. This proves the first part of the corollary. The
+
property of D̄ls follows similarly. 
The corollary shows that it may be necessary to introduce additional
RHP open-loop poles and zeros for decoupling. However these will not occur
at completely new positions; only the multiplicity of individual poles and
zeros of the plant might increase up to a factor m − 1. The corollary also
shows that no additional unstable poles and zeros need to be introduced for
decoupling if and only if Ng+ and Dg+ are both block diagonal. Note that this
condition means that G+ U is block diagonal for some unimodular matrix U ,
where G+ is an arbitrary matrix taken from a stability factorization.

Example 5.3.5. Consider the plant of Example 5.3.3. For the partition (2, 1),
it has been shown that (PO) is solvable. In addition, one can check that G+ U
+
is block diagonal for the unimodular matrix U = (Dls (2, 1))−1 . Moreover, the
controller K1 given in Example 5.3.3 is free of unstable poles and zeros, as
expected. ♦

Using the arguments of Doyle and Stein (1981), we intuitively expect that
no achievable performance is lost by decoupling if Ng+ and Dg+ are both block
diagonal. This conjecture is verified in the following section for a large class
of performance measures.

5.3.3 Achievable Performance after Decoupling

In this section we assume that the plant G and partition (mi ) are such that
the solvability conditions of Theorem 5.3.1 are satisfied. We study to what
extend the remaining free parameters, as specified in Theorem 5.3.2, can be
used to satisfy additional specifications. We determine a class of plants for
which the achievable performance after decoupling is not inferior to the one
in not necessarily decoupled systems.
It is well known (Freudenberg and Looze, 1988) that many performance
specifications, such as disturbance rejection, tracking and robustness can be
formulated in terms of S = (I +GK)−1 and T = GK(I +GK)−1 . We consider
performance measures J of the form
 
S
J( ) : R2m×m (s) → R.
T
158 5. Polynomial Matrix Approach

satisfying
   
S d(S)
J( ) ≥ J( ), (5.41)
T d(T )

where the operation d(M ) extracts the diagonal blocks of size mi × mi

from a matrix M , i.e., d(M ) := block diag{Mii }. The inequality (5.41) is
satisfied for most measures of practical interest. Consider for instance the
measure Jp defined by the Hp -norm of [(W1 SW2 )T (W3 T W4 )T ]T , where
W1 , W2 , W3 , W4 ∈ Rm×m
d (s) are stable weighting matrices. Dickman and
Sivan (1985) have shown that J∞ satisfies (5.41). For other performance
measures, such as J2 , (5.41) is also easily verified. Let J ∗ be the optimal
performance, i.e., the infimum of J over all stabilizing controllers, and Jd∗ be
the optimal decoupled performance, i.e., the infimum of J over all stabilizing
and decoupling controllers. Then Jd∗ ≥ J ∗ , and no achievable performance
(as measured by J) is lost by decoupling if Jd∗ is equal to J ∗ . The results
of this section will imply that this is the case provided the matrix G+ in a
stability factorization is block diagonal.
Some further notation is required. Let T (Td ) be the set of complimentary
sensitivity functions achievable by stabilizing (stabilizing and diagonalizing)
controllers, i.e.,

T := {T = GK(I + GK)−1 : K internally stabilizes G}

and

Td := {T ∈ T : T ∈ Rm×m
d (s)}.

Note that the elements of Td have been characterized in Theorem 5.3.2.

Moreover, let

d(T) := {d(T ) : T ∈ T, d(T ) nonsingular}.

In the definition of the above sets we allow the controllers to be non-proper.

This is done in order to simplify the development and the notation. If the
controllers are constrained to be proper, then zeros and poles at infinity have
to be considered in addition to finite ones. It is obvious that Td ⊂ d(T).
The following theorem gives a sufficient (and almost necessary) condition for
Td = d(T) to hold. The proof is deferred to the end of this section.

Theorem 5.3.3. The equality Td = d(T) holds if the polynomial matrices

Ng+ and Dg+ , as defined in Corollary 5.1.1, are both block diagonal, i.e.,

Ng+ ∈ Rm×m
d [s], Dg+ ∈ Rm×m
d [s]. (5.42)

Suppose in addition that det(N + ) and det(D+ ) are coprime. Then Td is equal
to d(T) if and only if (5.42) holds.
 5.3 Block Decoupling for General Plants 159

Note that (5.42) holds if and only if there exists a stability factorization
with a block diagonal G+ . Moreover, by Corollary 5.3.1, (5.42) holds if and
only if (PO) can be (easily) solved by the stable and minimum phase con-
troller K1 = (G− )−1 . However, this does not mean that (P) can be solved
by a stable and minimum phase controller, because stabilization might re-
quire unstable poles and zeros in Kd (Vidyasagar, 1985). For the special case
that the plant is already block diagonal, Theorem 5.3.3 means that the cross-
coupling terms in the controller are of no use for performance optimization
with respect to J.
The assumption on the determinants is quite weak; it is satisfied for an
open and dense set of plants. Nevertheless, at least from a theoretical point
of view, it is of interest to analyze if this assumption is really required to
prove the necessity of (5.42). We conjecture that this is not the case.
For a measure J satisfying (5.41), Td = d(T) implies Jd∗ = J ∗ . This
proves the following corollary.
Corollary 5.3.2. Let the plant satisfy (5.42) and the performance measure
satisfy (5.41). Then there are no performance limitations imposed by decou-
pling, in the sense that Jd∗ = J ∗ holds.
Youla, Jabr and Bongiorno (1976) have shown that the optimal controller
of the H2 - problem leads to a decoupled system provided that the plant
is square, stable and minimum phase. Corollary 5.3.2 thus generalizes their
result to possibly unstable, non-minimum-phase, or nonsquare plants, to the
block decoupling problem, and to a wide class of measures including J∞ and
J2 .
Example 5.3.6. Consider the plant of Example 5.3.1. We have already shown
that (P) is solvable if m1 = 2 and m2 = 1. The matrix N + in Example 5.3.3
is already in Hermite form, so that Ng+ = N + . The Hermite form of D+ is
given by
 
1 1 0
Dg+ =  0 s − 1 0  ,
0 0 1
which is block diagonal with respect to the partition m1 = 2 and m2 = 1.
Thus we will not lose any H∞ - and H2 - performance (in the sense of J∞ and
J2 ) by designing separate optimal controllers Kd,1 and Kd,2 for
 −1

+ 1 s−1
G1 =
0 1

and G+
2 = s − 1, respectively, and then using
 
Kd,1 0
−1 
K = (G−
g ) 0 Kd,2 
0 0
160 5. Polynomial Matrix Approach

as a controller for G, where [Ng+ (Dg+ )−1 0]G− g = G is a stability factorization.

Recall that, in this subsection, we allow the controller to be non-proper. A
proper controller for this example is obtained by choosing stable polynomial
matrices P1 and P2 such that (G− g )
−1
(block diag{P1 , P2 })−1 is proper, and
designing proper optimal controllers Kd,1 and Kd,2 for the strictly proper
systems G+ 1 P1
−1
and G+ −1
2 P2 , respectively, compare (5.36). ♦
Note that the condition (5.42) is far from being necessary for the H∞
case. This is in contrast to the condition in Theorem 5.3.3 and comes from
the fact that H∞ -norm of a block diagonal matrix is equal to the maximum of
the norms of the diagonal blocks. Dickman and Sivan (1985) give an example
illustrating this effect.

Proof of Theorem 5.3.3. Sufficiency: Suppose that N + and D+ are both

block diagonal. It suffices to prove the claim for v = 2. The general case then
follows by using induction on v. Consider any stabilizing controller K for G
and T = GK(I + GK)−1 . We have to construct another stabilizing controller
K̄ for G such that T̄ := GK̄(I + GK̄)−1 = d(T ). The matrix G+ is block
diagonal, i.e., G+ = block diag{G+ +
1 , G2 }. Since K stabilizes G, it follows that
K̃ := [I 0]G− K stabilizes G+ . Consider now Figure 5.2. Internal stability
means (Chapter 3) that the four transfer function matrices r → z, r → y,
d → z, and d → y, are stable. In particular, the transfer function matrices,
r1 → z1 , r1 → y1 , d1 → z1 and d1 → y1 , are stable. This implies that the
encircled controller Kd,1 internally stabilizes G+ 1 . Its transfer function is

Kd,1 = K̃11 + K̃12 (I + K̃22 G+

2)
−1
,
where K̃ij ∈ Rmi ×mj (s) is the (i, j) block of K̃, i, j = 1, 2. Similar arguments
for the second channel show that
Kd,2 = K̃22 + K̃21 (I + K̃11 G+
1)
−1

internally stabilizes G+ +
2 . Hence, Kd := block diag{Kd,1 , Kd,2 } stabilizes G ,
− −1 T + −
implying that K̄ := (G ) [I 0] Kd stabilizes G = [G 0]G . Moreover,
we have T̄ := GK̄(I + GK̄)−1 = G+ Kd (I + G+ Kd )−1 = d(T ).
Necessity: Assume first that Ng+ is not block diagonal. Let Ni+ ∈
mi ×m
R [s] be the i-th block-row of Ng+ , Li be the greatest common left divisor
of the columns of Ni+ , and L := block diag{Li }. Then, Ng+ = LÑ + for some
polynomial matrix Ñ + , which can be partitioned as follows:
 +
Ñ1
 Ñ + 
 2 
Ñ + =  .  , Ñi+ = [0 Ni Mi ],
 .. 
Ñv+
where
Pv Ñi+ ∈ Rmi ×m [s], Ni ∈ Rmi ×mi [s], and Mi ∈ Rmi ×pi [s]; pi :=
j=i+1 mj . Let k (1 ≤ k < v) be the index such that Mk 6= 0 and Mi = 0
 5.3 Block Decoupling for General Plants 161

Kd,1
H d1 d2
H
H
HH
j
r1 z1 y1
-d - - d? - G+ q -
2
– 6
K̃
r2 z2 y2
-d - - d?
- G+ q -
2
– 6

Fig. 5.2. Controller construction

for i > k. Since the degrees of the off-diagonal elements of a row in Ng+ are
strictly smaller than the degree of the corresponding diagonal element, the
matrix Nk has at least one zero s0 ∈ C+ . For this zero there exist y0 ∈ Cmi
and z0 ∈ Cpi such that
y0T Nk (s0 ) = 0,
y0T Mk (s0 ) 6= 0,
[0, y0T , z0T ]Ñ + (s0 ) = 0.
+
Consider now a right strict adjoint Nra of N + . Since Ñ + Nra
+
is block diagonal,
the latter equality implies

y0T Ñk+ (s0 )Nra

+
(s0 ) = 0.

Theorem 5.3.2 shows that, for any stabilizing and block decoupling controller
K̄ for G, we have GK̄ = Qd Kd , where Qd = N + Nra +
(D̄+ D̄+ )−1 and Kd
stabilizes Qd . The parameterization of all stabilizing controllers Kd for Qd
(Chapter 3) implies that each element T̄ in Td has the form

T̄ = N + Nra
+
H̄d ,

where H̄d is some block diagonal stable rational matrix. Its k-th diagonal
block T̄k becomes T̄k = Lk P̄k , where P̄k = Ñk+ Nra +
H̄d [0 I 0]T satisfies
T
y0 P̄k (s0 ) = 0.
We now construct a particular stabilizing controller K for G such that the
k-th diagonal block Tk of T = GK(I + GK)−1 is of the form Tk = Lk Pk with
y0T Pk (s0 ) 6= 0. This then implies that d(T ) ∈ d(T) but d(T ) 6∈ Td , which will
complete the proof.
Let G = N D−1 be a right coprime polynomial matrix fraction of G with
N = N + N − for some stable polynomial matrix N − . It follows again from the
162 5. Polynomial Matrix Approach

parameterization of all stabilizing controllers for G that T is parameterized

by
T = N (−Ỹ + QD̃),
where Q runs through all stable rational matrices, and the polynomial ma-
trices Ỹ and D̃ satisfy the Bezout identity (Chapter 2). The k-th diagonal
block of T has the form Tk = Lk Pk , where Pk = Ñk+ N − (−Ỹ + QD̃)[0 I 0]T .
Note that N − (s0 ) is nonsingular, so that y0T Ñk+ (s0 )N − (s0 ) 6= 0. We have
det(D̃) = det(D), so that D̃(s0 ) is nonsingular, by assumption. Hence
(−Ỹ (s0 ) + Q(s0 )D̃(s0 ))[0 I 0]T can be set to be an arbitrary matrix by a
proper choice of Q(s0 ). Especially we can choose Q such that y0T Pk (s0 ) 6= 0.
This Q defines the required controller.
The necessity of the other condition (D+ is block diagonal) can be proven
using similar arguments and appropriate dual concepts. 
In this section, firstly, the block decoupling problem with internal stabil-
ity by unity output feedback is constructively solved for possibly non-square
plants of full row rank. Necessary and sufficient solvability conditions (Theo-
rem 5.3.1) and a characterization of all controllers solving the problem (The-
orem 5.3.2) are given. Secondly, a class of systems is specified for which there
is no need to scarify achievable performance by putting an decoupling prec-
ompensator into the feedback loop (Theorem 5.3.3). These results together
can be used to determine a suitable grouping of the outputs for decoupling,
and to support the decision on whether to do the decoupling inside or outside
the feedback loop.

5.4 A Unified Solution

The approach to the general decoupling problem (P), which we have taken in
the previous two sections, is first treating the square case and then extending
to the non-square case. This is of course a natural way of research evolution
and facilitates understanding. It is however possible and also advantageous
to unify both cases and develop a solution in a general setting. Thus, in this
section, we will give the unified approach to (P), which is also independent of
Sections 5.2 and 5.3. It may give neater results in some places (for example,
you may compare the parameterization of all the solutions later). The two
new concepts, least block multiples of a full rank polynomial matrix and
skew prime polynomial matrices in C+ , are introduced in resolution of the
problem.
To unify square and non-square cases, we need to modify the stability
factorization, G = [G+ 0]G− . But [G+ 0] is to be replaced by a single
matrix and for notational clarity it is still be denoted by G+ .

Definition 5.1.10 . Let G ∈ Rm×l (s) have full row rank, and assume that
there are a full row rank G+ ∈ Rm×l (s) and a nonsingular G− ∈ Rl×l (s)
 5.4 A Unified Solution 163

such that

G = G+ G− . (5.43)

(5.43) is called a stability factorization of G if the zeros and poles of G+

(respectively, G− ) are precisely the RHP (respectively, LHP) zeros and poles
(counting multiplicities) of G. Such a G− is then stable, minimum phase, and
invertible. G+ and G− in a stability factorization are here called the greatest
C + -divisor and the greatest C − -divisor of G, respectively.

It follows from Lemma 5.1.1 that this new stability factorization exists for
any full rank G(s). As a convention, the stability factorization in this section
is always referred to that in Definition 5.1.1 0 above.
For a full row rank polynomial matrix P ∈ Rm×l [s], denote by ∆(P ) the
m-th determinantal, the monic greatest common divisor of all m × m minors
of P (s). In terms of the Smith form, P = U [Λ 0]V with U and V unimodular
and Λ diagonal, there holds ∆(P ) = det(Λ). For a square P , ∆(P ) equals
det(P ) up to a nonzero constant. Let P = P + P − be a stability factorization
as in Definition 5.1.10 , define ∆+ (P ) := ∆(P + ), deg(P ) := deg(∆(P )), where
the right hand side of the last equality means the degree of the polynomial
∆(P ), and deg + (P ) = deg(P + ) = deg(∆(P + )). P is called antistable if
∆(P ) = ∆(P + ), that is, P has all its zeros in C+ .
Let P ∈ Rm×l [s] be a full row rank polynomial matrix. The representa-
tion, P = LU , is said to be an LU-factorization of P if L is a nonsingular
polynomial matrix and U is an irreducible polynomial matrix, i.e., U (s) has
full row rank for all s ∈ C.
Lemma 5.4.1. For an LU-factorization, P = LU , there holds ∆(P ) =
∆(L). In general, let P ∈ Rm×l [s] and P1 ∈ Rm×k [s] be both of full row
rank with k ≥ m such that P = P1 P2 for some polynomial matrix P2 , then
∆(P1 ) divides ∆(P ).
Proof. P has a Smith factorization, P = U1 [Λ 0]U2 , where U1 and U2 are
unimodular with Λ = diag{pi }. By definition, ∆(P ) = ∆(Λ). Express P
as P = LU , where L = U1 Λ and U = [I 0]U2 , then P = LU is an LU-
factorization and ∆(L) = ∆(Λ) = ∆(P ). Now, let P = L̃Ũ be another
LU-factorization, then LU = L̃Ũ . If L−1 L̃ = U3 is unimodular, there will
hold ∆(L̃) = ∆(L) = ∆(P ), which then completes the proof. Let AB −1 be a
coprime polynomial matrix fraction of U3 , then U = AB −1 Ũ . Lemma 5.1.2
says that Ũ has B as its left divisor but Ũ is irreducible and B is thus uni-
modular. Consider dually Ũ = BA−1 U , one concludes that A is unimodular,
too. The second statement follows from applying the Binet-Cauchy formula
(Gantmacher,1959) to the equality, P = L1 P̃ , where P1 = L1 U1 is an LU-
factorization and P̃ = U1 P2 . 
A strange thing about nonsquare polynomial matrices is that an anti-
stable P can have a stable divisor for itself. For instance, we have [s 0] =
164 5. Polynomial Matrix Approach

[s 0]diag{1, s + 1}. However, the following lemma shows that such a stable
divisor has certain structure.

Lemma 5.4.2. Let P ∈ Rm×l [s] and P1 ∈ Rm×l [s] be both of full row rank
such that P = P1 P2 for some stable and nonsingular polynomial matrix P2 .
If P is antistable, then U P2−1 is a polynomial matrix, where P = LU is an
LU-factorization.

Proof. Assume conversely that it is not the case, let P̃2−1 Ũ be a left coprime
polynomial matrix fraction of U P2−1 with P̃2 stable and deg(P̃2 ) nonzero,
it then follows that LP̃2−1 Ũ = P1 . By Lemma 5.1.2, LP̃2−1 is a polynomial
matrix. This means that L has a stable divisor P̃2 which is not unimodular,
contradicting the fact that P is antistable. 

5.4.1 Skew Primeness in C+

The matrices D ∈ Rm×m [s] and N ∈ Rm×l [s] are called left coprime in C+ ,
if

Rank[D(s) N (s)] = m, f or all s ∈ C+ .

The notion of skew primeness needs to be generized, too.

Definition 5.4.1. Let a full row rank N ∈ Rm×l [s] and a nonsingular D ∈
Rm×m [s] be given. N and D are said to be externally skew prime in C+ if
there are a full row rank N̄ ∈ Rm×l [s] and a nonsingular D̄ ∈ Rl×l [s] such
that

DN = N̄ D̄ (5.44)

with D and N̄ left coprime in C+ and N and D̄ right coprime in C+ . Such

(N̄ , D̄) is refereed to as an internally skew prime pair of (N, D) in C+ .

Note that the skew primeness in C+ and ordinary skew primeness (in C) as
defined by Wolovich (1978) differ only in relevant domains, and all tests and
algorithms for ordinary skew prime polynomial matrices in Wolovich (1978)
and Wang,Sun and Zhou (1989) thus apply to the case of skew primeness in
C+ with the obvious modification of replacing C by C+ . Our development
does however require skew primeness in C+ . The following lemma shows that,
under certain conditions, skew primeness in C+ is equivalent to the ordinary
skew primness.

Lemma 5.4.3. Let a full row rank N ∈ Rm×l [s] and a nonsingular D ∈
Rm×m [s] be both antistable, then N and D are externally skew prime in C+
if and only if they are externally skew prime (in C).
 5.4 A Unified Solution 165

Proof. Assume that N and D are externally skew prime in C+ , then there are
internally skew prime N̄ and D̄ in C+ . By performing a stability factorization
for D̄T , we get D̄ = D̄− D̄+ with D̄− stable and D̄+ antistable. (5.44) then
becomes

DN = Ñ D̃,

where Ñ = N̄ D̄− and D̃ = D̄+ . The left coprimeness of D and N̄ in C+

implies that

Rank[D(s) Ñ (s)] = m, f or all s ∈ C+ , (5.45)

since D̄− (s) is nonsingular for all s ∈ C+ . Additionally, because D is anti-

stable, we also have

Rank[D(s) Ñ (s)] = m, f or all s ∈ C− . (5.46)

(5.45) and (5.46) together mean that D and Ñ are left coprime. In a similar
way, one can show that N and D̃ are right coprime, too. Thus, it follows that
N and D are externally skew prime. The converse implication is trivial. The
proof is completed. 
As one will see later, our solvability condition and compensator parame-
terization for (P) deal only with antistable N and D. Then, due to Lemma
5.4.3, a natural question which one is likely to ask is why the skew primeness
in C+ is here introduced. The real reason is that in spite of the equivalence
there do exist differences between the ordinary internally skew prime pair
(Ñ , D̃) and the internally skew prime pair ( N̄ , D̄) in C+ , as seen in the above
proof and will be further exhibited by the following example. Even more im-
portantly, it will turn out that a general compensator which solves (P) is
constructed from (N̄ , D̄) but not from (Ñ , D̃), and it therefore proves to be
necessary to introduce skew prime polynomial matrices in C+ .

Example 5.4.1. Let N = [s 0] and D = (s − 1). If we set Ñ = [s 0] and

D̃ = diag{(s − 1), 1}, then

DN = (s − 1)[s 0] = [s 0]diag{(s − 1), 1} = Ñ D̃.

One also sees that D and Ñ are left coprime and N and D̃ are right coprime.
(Ñ , D̃) is thus an internally skew prime pair of (N, D). Further, for a pair
(Ñ , D̃) to be an internally skew prime one of (N, D), it must have certain
properties. Take D̃ for comparison, it satisfies

∆(D̃) = ∆(D) = (s − 1). (5.47)

On the other hand, it can be easily verified that N̄ = [s 0] and D̄ = diag{(s−

1), p} with p being a stable polynomial constitute an internally skew prime
pair of (N, D) in C+ (but not in C) and D̄ satisfies
166 5. Polynomial Matrix Approach

∆(D̄) = (s − 1)p. (5.48)

(5.47) and (5.48) clearly exhibit an essential difference between ordinary skew
prime pairs and skew prime pairs in C+ for the same pair of externally skew
prime polynomial matrices N and D even when they are both antistable. ♦

For an ordinary skew prime pair (Ñ , D̃) of (N, D), we have that ∆(D) =
∆(D̃) and ∆(N ) = ∆(Ñ ) since N D̃−1 = D−1 Ñ with both fractions coprime.
In general, this does not hold for a skew prime pair in C+ . The following
lemma gives a test for skew primeness in C+ and characterizes zero poly-
nomials of an internally skew prime pair in C+ in terms of those of a given
externally skew pair.

Lemma 5.4.4. Let a full row rank N ∈ Rm×l [s] and a nonsingular D ∈
Rm×m [s] be both antistable, then
(i) N and D are externally skew prime in C+ if and only if there are N̄
and D̄ such that (5.44) holds true with ∆+ (D̄) = ∆(D) and N and D̄ right
coprime in C+ ;
(ii) for an internally skew prime pair (N̄ , D̄) in C+ of (N, D), there holds
the equality ∆(N̄ ) = ∆(N ); and
(iii) the equality ∆(D̄) = ∆(D) also holds if N is square. In this case,
there is no difference between ordinary internally skew prime pairs and in-
ternally skew prime pairs in C+ .

Proof. (i): For the sufficiency, we need only to show that D and N̄ are left
coprime in C+ . This is true if there holds

deg(D) = deg + (D1 ), (5.49)

where D1−1 N̄1 is a coprime polynomial matrix fraction of D−1 N̄ . Let N1 (D̄1 )−1
be a coprime polynomial matrix fraction of N (D̄)−1 . As N (D̄)−1 has been
assumed to be coprime in C+ , there holds

deg + (D̄1 ) = deg + (D̄). (5.50)

(5.44) can be rewritten as N1 (D̄1 )−1 = (D1 )−1 N̄1 with both sides being
coprime, implying that

deg + (D̄1 ) = deg + (D1 ). (5.51)

Collecting (5.50), (5.51), and the assumed condition, deg + (D̄) = deg(D),
gives (5.49).
For necessity, assume now that antistable matrices N and D are externally
skew prime and a dual internally skew prime polynomial matrices in C+ are
N̄ and D̄. Let R be a greatest common right divisor of N and D̄ so that

D̄ = D̄1 R, and N = N1 R. (5.52)

 5.4 A Unified Solution 167

Then, (5.44) becomes

DN1 = N̄ D̄1 , (5.53)

where N1 D̄1−1 is a right coprime. Moreover, D and N̄ are coprime since they
are coprime in C+ and D is antistable. Therefore, (5.53) means that ( N̄ , D̄1 )
is an internally skew prime pair of (N1 , D) and we thus have

∆(D̄1 ) = ∆(D), (5.54)

which implies that D̄1 is antistable, too. Since N and D̄ are coprime in C+ ,
their common divisor R must be stable and ∆+ (D̄) = ∆+ (D̄1 R) = ∆+ (D̄1 ) =
∆(D̄1 ). (5.54) then becomes ∆+ (D̄) = ∆(D).
(ii): It follows also from (5.53) that

∆(N̄ ) = ∆(N1 ). (5.55)

By Lemma 5.4.1, (5.52) implies that

∆(N1 ) divides ∆(N ). (5.56)

By Lemma 5.4.2, (5.52) can be rewritten as LP = N1 , where N = LU is

a LU-factorization and P = U R−1 is a polynomial matrix. It then follows
again from Lemma 5.4.1 that

∆(L) = ∆(N ) divides ∆(N1 ). (5.57)

(5.55)-(5.57) together give the required result.

(iii): If N is square, then (5.52) implies that ∆(R) divides ∆(N ) and R
is antistable. But it has been shown in (i) that R is stable and it is thus
unimodular. This, together with (5.54), implies that

∆(D̄) = ∆(D̄1 R) = ∆(D̄1 ) = ∆(D).

The proof is completed. 

5.4.2 Least Block Multiples

The notion of strict adjoints, as introduced by Hammer and Khargonekar

(1984) played an important role in the solution of diagonal decoupling prob-
lem (Section 2). Recall that for a nonsingular matrix P ∈ Rm×m [s], a polyno-
mial matrix Pra is said to be a right strict adjoint of P if P Pra ∈ Rm×m
d [s] and
Pra is a left divisor of any polynomial matrix Pa for which P Pa ∈ Rm×m d [s],
where the partition is mi = 1, i = 1, 2, · · · , v. The following example suggests
that this notion can not be extended to the nonsquare case, and motivates the
alternative concept of least block multiples, which will be formally introduced
in Definition 5.4.2 below.
168 5. Polynomial Matrix Approach

Example 5.4.2. Consider a polynomial matrix be P = [s − 1 s]. If we take

P1 = [1 0]T and P2 = [0 1]T , then both P P1 and P P2 are diagonal and
nonsingular. However, there exists no polynomial matrix Pra such that P Pra
is nonzero and Pra a common divisor of both P1 and P2 . ♦
One can however consider products P Pi instead of Pi themselves. It is eas-
ily verified that P P1 and P P2 are both multiples of P P3 with P3 = [−1 1]T .
It turns out that such a P P3 can always be constructed so that the following
definition makes sense.
Definition 5.4.2. Let P ∈ Rm×l [s] have full row rank, a block diagonal
and nonsingular polynomial matrix Pm ∈ Rm×m d [s] is said to be a right block
multiple of P with respect to the partition {mi } if there is a polynomial matrix
Pa ∈ Rl×m [s] such that Pm = P Pa . A right block multiple Prm is further said
to be a least right block multiple of P if for any right block multiple Pm
of P with respect to the same partition, Prm is a left divisor of Pm , i.e.,
Pm = Prm Pd for some Pd ∈ Rm×m d [s].
In Example 5.4.2, P P3 = 1 is a least right block multiple. In general, we
need a procedure for constructing a least right block multiple in a systematic
way. Given a full row rank P ∈ Rm×l [s] and a partition {mi }, we express P
as an LU-factorization
P = LU,
where L ∈ Rm×m [s] is nonsingular and U ∈ Rm×l [s] is irreducible so that
there is a polynomial matrix X such that
LX = I.
L−1 is then partitioned as
L−1 = [L̃1 L̃2 · · · L̃v ],
where L̃i ∈ Rm×mi (s), i = 1, 2, · · · , v. We further factorize each L̃i as a right
coprime polynomial matrix fraction
L̃i = Ai Bi−1 , i = 1, 2, · · · , v,
where Ai ∈ Rm×mi [s] and Bi ∈ Rmi ×mi [s]. Then, L−1 becomes
L−1 = AB −1 ,
where
A = [A1 A2 · · · Av ],
and
B = block diag{B1 B2 · · · Bv }. (5.58)
The following lemma shows that B is just a least right block multiple.
 5.4 A Unified Solution 169

Lemma 5.4.5. Let P be a full row rank polynomial matrix, then P has a
least right block multiple with respect to every partition. A particular least
right block multiples of P is given by Prm = B, as defined in (5.58). Further,
for a fixed partition, each diagonal block in a least right block multiple of P is
right unimodular equivalent to the corresponding diagonal block in all other
least right block multiples of P .

Proof. Take Pra = XA ∈ Rl×m [s], it then follows that P Pra = P XA =

LA = B so that B is a right block multiple of P . Assume now that Pm be
an arbitrary right block multiple of P . This means that Pm = P Pa for some
polynomial matrix Pa , equivalently,

U Pa = L−1 Pm = [A1 B1−1 Pm11 A2 B2−1 Pm22 · · · Av Bv−1 Pmvv ],

where Pm = block diag{Pmii }. One sees from the above equation that
Ai Bi−1 Pmii are polynomial matrices, but as Ai Bi−1 are coprime, Bi−1 Pmii
are thus polynomial matrices, or Pm = BPd for some block diagonal polyno-
mial matrix Pd . Therefore, B is a left divisor of any right block multiple of P
0
and by definition it is a least right block multiple of P . Let Prm and Prm be
both least right block multiples of P with respect to the same partition, it fol-
0 0
lows from the definition that Prm = Prm P1 and Prm = Prm P2 for some block
diagonal polynomial matrices P1 and P2 , implying that Prm = Prm P2 P1 and
both P1 and P2 are unimodular. Since they are block diagonal, all the di-
agonal blocks of P1 and P2 are unimodular, too. The proof is completed.


Example 5.4.3. Consider a 3 × 4 polynomial matrix

 
1 1 0 2
P =  1 s 0 s + 1 .
0 s 1 s+1

An LU-factorization is given by P = LU with

 
1 1 0
L =  1 s 0 .
0 s 1

Its inverse is
 s −1

s−1 s−1 0
 
L−1 = 

−1
s−1
1
s−1 0 
.
s −s
s−1 s−1 1

Let first the partition be m1 = 2 and m2 = 1, then the corresponding sub-

matrices L̃1 and L̃2 of L−1 are of sizes 3 × 2 and 3 × 1, respectively. Their
coprime polynomial matrix fractions are constructed as follows
170 5. Polynomial Matrix Approach

 
s −1  

s−1 s−1
 1 −1  −1
L̃1 =  −1 1 = 0 1  1 0
= A1 B1−1 ,
 s−1 s−1  1 s−1
s −s 0 −s
s−1 s−1

and
   
0 0
L̃2 =  0  =  0  [1]−1 = A2 B2−1 .
1 1

We then obtain a least right block multiple Prm (2, 1) with respect to the
partition {2, 1} as
 
1 0 0
Prm (2, 1) =  1 s − 1 0  .
0 0 1

On the other hand, if a different partition, m1 = 1, m2 = 2, is considered,

then the relevant submatrices and their coprime fractions become
 s 
 
s−1
 −1  s
L̃1 =     −1
 s−1  = −1 (s − 1) = A1 B1 ,
−1

s s
s−1

and
 
−1
0  

s−1
 −1 0  −1
L̃2 =  1
0  =  1 0  s−1 0 = A2 B2−1 .
 s−1  0 1
−s −s 1
s−1 1

A least right block multiple Prm (1, 2) with respect to the partition {1, 2} is
given by
 
s−1 0 0
Prm (1, 2) =  0 s − 1 0 . ♦
0 0 1

We are now interested in the zero property of least right block multiples.
In Example 5.4.3, there hold ∆(P ) = (s − 1), ∆(Prm (2, 1)) = (s − 1), and
∆(Prm (1, 2)) = (s − 1)2 . The following lemma relates the zero polynomial of
Prm to that of P .

Lemma 5.4.6. Let Prm be a least right block multiple of a full row rank poly-
nomial matrix P , then, ∆(P ) divides ∆(Prm ) and ∆(Prm ) divides (∆(P ))m .
 5.4 A Unified Solution 171

Proof. By definition, for any least right block multiple Prm there is a poly-
nomial matrix Pra such that P Pra = Prm , which implies from Lemma
5.4.1 that ∆(P ) divides ∆(Prm ). Let P = LU be an LU-factorization and
X a polynomial matrix such that U X = I. Take Pa = X adj(L), where
adj(L) is the adjoint of L, that is, L adj(L) = det(L) I, It then follows
that det(L) I is a right block multiple of P with respect to any partition
as P Pa = LU X adj(L) = det(L) I. Therefore, there is a block polynomial
matrix P1 such that det(L) I = Prm P1 . By Lemma 5.4.1, ∆(Prm ) divides
∆(det(L) · Im ) = (∆(L))m = (∆(P ))m , noting that det(L) differs from ∆(L)
only by a non-zero constant. Hence, the result. 
Lemma 5.4.6 shows that Prm does not have any zeros different from those
of P . Especially, Prm will be stable (respectively, antistable) if P has this
property. Example 5.4.3 demonstrates that multiplicity of a zero of Prm may
be greater than the corresponding multiplicity in P .
If the diagonal partition is taken into account, that is, mi = 1, i =
1, 2, · · · , v, v = m, one can expect a natural connection between least right
block multiples and strict right adjoints. The following corollary serves this
purpose.

Corollary 5.4.1. For a nonsingular polynomial matrix P , Pra is a right

strict adjoint of P if and only if P Pra is a least right block multiple of P with
respect to the partition, m1 = m2 = · · · = mv = 1.

Proof. If Pra is a strict right adjoint of P , then P Pra ∈ Rm×m

d [s] and for any
Pa such that P Pa = Pm is diagonal and nonsingular, there holds Pa = Pra P1
for some diagonal polynomial matrix P1 . It also means that Pm = P Pra P1
and Prm = P Pra is thus a least right diagonal multiple of P . Conversely,
assume that Prm is a least right diagonal multiple of P , then P Pra = Prm for
some polynomial matrix Pra . Furthermore, for any Pa such that P Pa = Pm is
diagonal and nonsingular, we have Pm = Prm P1 for some diagonal polynomial
matrix P1 . This implies that P Pa = P Pra P1 , i.e., Pa = Pra P1 , and then Pra
is a strict right adjoint of P . The proof is completed. 
Let Pa be a polynomial matrix such that P Pa = Pm be a right block
multiple of a full row rank polynomial matrix P . Then, in general, Pa can
not be expressed as Pa = Pra P1 for any Pra such that P Pra = Prm is
a least right block multiple of P , as shown by the following example. Let
P = [1 s], set Pa = [0 1]T , then P Pa = s. It is obvious that Prm = 1. Take
∗
Pra = [1 0]T and Z = [s − 1]T , then a general solution of the equation
∗
P Pra = Prm for Pra is Pra = Pra + Zw = [(1 + sw) w]T . One can see that
Pa = Pra p will never hold for any polynomials w and p. In fact, Pa can be
∗
only expressed as Pa = Pra p + Zw with p = s and w = −1. This is why
least block multiples are introduced and the numerator polynomial matrix of
a general decoupling compensator has the form in (5.94), which is much less
+
nice than N̄k = N̄ra N̄q1 if the latter were true.
172 5. Polynomial Matrix Approach

It is obvious that the concept of least left block multiples for a full column
rank polynomial matrix can be defined in a dual way and the dual results to
least right block multiples all hold.

5.4.3 Solvability Conditions

Let us consider (P) now. Combining Theorem 5.1 with the stability factor-
ization in (5.43), we can obtain the following simplification.
Theorem 5.4.1. For any full row rank plant G ∈ Rm×l p (s), there is a K
such that GK is block diagonal and nonsingular, and has no RHP pole-zero
cancellations if and only if there is a K + which can do so with G replaced by
G+ , where G = G+ G− is a stability factorization.
Proof. For a stability factorization G = G+ G− , we have

δ + (G) = δ + (G+ ). (5.59)

Let K and K + be related by

K + = G− K, or, K = (G− )−1 K + ,

then, as G− is nonsingular, stable, and minimum phase, K and K + satisfy

δ + (K) = δ + (K + ). (5.60)

GK is nonsingular and block diagonal if and only if so is G+ K + since GK =

G+ G− K = G+ K + , which also means that

δ + (GK) = δ + (G+ K + ). (5.61)

It follows from (5.59)-(5.61) that

δ + (GK) = δ + (G) + δ + (K)

if and only if

δ + (G+ K + ) = δ + (G+ ) + δ + (K + ).

The proof is completed. 

In view of Theorem 5.4.1, the greatest C− -divisor G− of a plant G has no
effect on the solvability of (P). Instead, we need only to consider the greatest
C+ -divisor of G+ . By Definition 5.1.10 , G+ ∈ Rm×l (s) has full row rank. We
further factorize G+ as

G+ = (D+ )−1 N + ,

where D+ ∈ Rm×m [s] is nonsingular, N + ∈ Rm×l [s] has full row rank, and
D+ and N + are both antistable. By a dual statement to Lemma 5.4.5, there
+
is a least left block multiple Dlm ∈ Rm×m
d [s] of D+ such that
 5.4 A Unified Solution 173

+ + +
Dla D = Dlm
+
for some nonsingular Dla ∈ Rm×m [s]. It follows from Lemmas 5.4.1 and 5.4.6
+ +
that Dla are antistable. Now, if N + and Dla are externally skew prime in
+
C+ , then there exist internally skew prime N̄ + ∈ Rm×l [s] and D̄la ∈ Rl×l [s]
in C+ such that
+ + +
Dla N = N̄ + D̄la
+
with N̄ + of full row rank and D̄la nonsingular. Let N̄rm+
be a least right
+
block multiple of N̄ . We are now in a position to state the main solvability
condition for (P).

Theorem 5.4.2. For a full row rank plant G, (P) is solvable if and only
+ +
if N + and Dla are externally skew prime in C+ and Dlm +
and N̄rm are left
coprime.

Proof: In the proof of necessity, we need the following lemmas.

Lemma 5.4.7. Let A, B, C, D, and N be all polynomial matrices such that

B −1 AC −1 = D−1 N with B −1 A and D−1 N both coprime, then B is a right
divisor of D, that is, D = D1 B for some polynomial matrix D1 .

Proof. Let C̃ −1 Ã be a left coprime polynomial matrix fraction of AC −1 , it

then follows (Chapter 2) that (C̃B)−1 Ã is left coprime and thus D = U C̃B
for some unimodular polynomial matrix U . Hence the result. 

Lemma 5.4.8. Let (D+ )−1 N + and Nk Dk−1 be, respectively, coprime poly-
nomial matrix fractions of G+ and K, G+ K have no unstable pole-zero can-
cellations, and D+ and N + be both antistable, then (D+ )−1 (N + Nk ) is left
coprime and (N + Nk )(Dk )−1 is right coprime in C+ .

Proof. Assume conversely that (D+ )−1 (N + Nk ) is not coprime, then a coprime
one (D̃+ )−1 Ñ of it will satisfy

deg (D̃+ ) < deg (D+ ).

We therefore have
δ + (G+ K) = δ + ((D̃+ )−1 Ñ Dk−1 )
≤ deg (D̃+ ) + deg + (Dk )
< deg (D+ ) + deg + (Dk )
= δ + (G+ ) + δ + (K),
which contradicts the assumption that G+ K has no unstable pole-zero cancel-
lations. (D+ )−1 (N + Nk ) is thus left coprime. The coprimeness of (N + Nk )(Dk )−1
in C+ follows similar lines. The proof is completed. 
174 5. Polynomial Matrix Approach

According to Theorem 5.4.1, we can assume, without loss of generality,

that the plant is such that G = G+ .
NECESSITY: Let Nk Dk−1 and Dq−1 Nq be coprime polynomial matrix frac-
tions of K and Q = G+ K, respectively. Then, we have

Q = Dq−1 Nq = (D+ )−1 N + Nk Dk−1 . (5.62)

If K solves (P) for G+ , it follows from Theorem 5.4.1 that K is such that
G+ K is block diagonal and has no unstable pole-zero cancellations. The latter
fact further implies by Lemma 5.4.8 that (D+ )−1 (N + Nk ) is coprime and
Lemma 5.4.7 then says that

Dq = D1 D + (5.63)

for some polynomial matrix D1 . Since Q is block diagonal, we can have a

coprime polynomial matrix fraction of it with its numerator and denominator
being both block diagonal, and assume that Dq−1 Nq is just such a fraction.
By a dual result to Lemma 5.4.5, (5.63) for a block diagonal Dq implies that
+ + +
Dq = Dq1 Dlm = Dq1 Dla D , (5.64)
+
where Dq1 is block diagonal polynomial matrix, Dlm is a least left block
+ +
multiple of D , and Dla is an antistable polynomial matrix such that
+ + +
Dla D = Dlm . Substituting (5.64) into (5.62) gives
+ +
Dla N Nk Dk−1 = N̄q D̄q1
−1
, (5.65)

where
−1 −1
N̄q D̄q1 = Dq1 Nq (5.66)

is block diagonal and polynomial matrices N̄q and D̄q1 are chosen such that
−1
they are both block diagonal and right coprime. Since N̄q D̄q1 is coprime and
(5.65) holds, D̄q1 is a left divisor of Dk , i.e.,

Dk = D̄q1 Dk1 , (5.67)

for some polynomial matrix Dk1 . Substituting (5.67) into (5.65) yields
+ + −1
Dla N D̄k1 N̄k = N̄q , (5.68)

where
−1 −1
D̄k1 N̄k = Nk Dk1 (5.69)

with the left hand side of (5.69) being a left coprime polynomial matrix
fraction. Note from (5.68) that this coprimeness, together with the fact that
N̄q is a polynomial matrix, implies, by Lemma 5.1.2, that
 5.4 A Unified Solution 175

+ + −1
Dla N D̄k1 = N̄ + (5.70)

is a polynomial matrix and of full row rank. One knows from Lemma 5.4.8
that no unstable pole-zero cancellations in G+ K implies that N + Nk and Dk
are right coprime in C+ and this also ensures that

N + and D̄k1 are right coprime in C+ (5.71)

because N + Nk Dk−1 = N + Nk Dk1 −1 −1 −1

D̄q1 = N + D̄k1 −1
N̄k D̄q1 and the involved
−1 −1 −1
fractions Nk Dk , Nk Dk1 , and D̄k1 N̄k are all coprime there. The detailed
proof of (5.71) is similar to that of Lemma 5.4.8 and is therefore omitted.
We now proceed to show that
+
∆(Dla ) = ∆+ (D̄k1 ). (5.72)

It follows from (5.64) and (5.66) that

+
∆+ (Dq ) = ∆+ (Dq1 )∆(Dla )∆(D+ ), (5.73)

and

∆+ (Dq1 ) = ∆+ (D̄q1 ). (5.74)

No unstable pole-zero cancellations in G+ K mean that

∆+ (Dq ) = ∆+ (Dk )∆(D+ ). (5.75)

Combining (5.73)-(5.75) gives

+
∆+ (Dk ) = ∆+ (D̄q1 )∆(Dla ). (5.76)

Further, (5.67) gives

∆+ (Dk ) = ∆+ (D̄q1 )∆+ (Dk1 ). (5.77)

Since both fractions in (5.69) are coprime, we also obtain

∆+ (Dk1 ) = ∆+ (D̄k1 ). (5.78)

Collecting (5.76)-(5.78) yields (5.72). By Lemma 5.4.4, (5.70)-(5.72) together

+
mean that N + and Dla are externally skew prime in C+ .
With (5.70), (5.68) becomes

N̄ + N̄k = N̄q . (5.79)

Since N̄q is block diagonal, it follows from Lemma 5.4.5 that there is some
N̄q1 ∈ Rm×m
d [s] such that
+
N̄rm N̄q1 = N̄q , (5.80)
176 5. Polynomial Matrix Approach

+
where N̄rm is a least right block multiple of N̄ + . From (5.64), (5.66), and
(5.80), Q = G+ K can be expressed as
+ −1 + −1 −1
Q = Dq−1 Nq = (Dq1 Dlm ) Nq = (Dlm ) Dq1 Nq
+ −1 −1 + −1 + −1
= (Dlm ) N̄q D̄q1 = (Dlm ) N̄rm N̄q1 D̄q1 . (5.81)
+ −1 +
Because Dq−1 Nq is coprime, so is (Dlm ) N̄rm . The proof is completed. 
SUFFICIENCY: If the conditions in the theorem are satisfied, there is an
+
internally skew prime pair (N̄ + , D̄la ) in C+ such that
+ + +
Dla N = N̄ + D̄la (5.82)
+
with N + and D̄la right coprime in C+ , and
+ +
∆(Dla ) = ∆+ (D̄la ). (5.83)
+
Let N̄rm be a least right block multiple of N̄ + such that
+
N̄ + N̄ra +
= N̄rm (5.84)
+
for some polynomial matrix N̄ra . We then take
+ −1 +
K1 = (D̄la ) N̄ra . (5.85)
as a candidate for the required compensator K. One sees that
δ + (G) = deg(D+ ), (5.86)
and
+
δ + (K1 ) ≤ deg + (D̄la ). (5.87)
By (5.82) and (5.84), the loop transfer matrix can be rewritten as
+ −1 +
Q1 = GK1 = (D+ )−1 N + (D̄la ) N̄ra
+ −1 + + + + −1 + +
= (D+ )−1 (Dla ) N̄ N̄ra = (Dla D ) N̄ N̄ra
+ −1 +
= (Dlm ) N̄rm , (5.88)
+
which is block decoupled. The assumed left coprimeness of Dlm and N̄rm
implies that
+ +
δ + (Q1 ) = deg(Dlm ) = deg(D+ ) + deg(Dla )
+ + +
= deg(D ) + deg (D̄la ), (5.89)
where the last equality follows from (5.83). Therefore, collecting (5.86), (5.87),
and (5.89) gives
δ + (Q1 ) ≥ δ + (G) + δ + (K1 ), (5.90)
which ensures that no unstable pole-zero cancellations will occur in GK1 .
Thus, by Theorem 5.4.1, (P) is solvable for G+ . The sufficiency is proven. 
Let us take an example for illustration.
 5.4 A Unified Solution 177

Example 5.4.4. Let a 3 × 4 plant be given by

 1 1 2

s+1 s+2 0 s+4
 
G= 1 s
 (s−1)(s+1) (s−1)(s+2) 0 s+1
(s−1)(s+4)
.

s 1 s+1
0 (s−1)(s+2) (s−1)(s+3) (s−1)(s+4)

A stability factorization of it is G = G+ G− , where G− = diag −1 {(s + 1), (s +

2), (s + 3), (s + 4)} and
 
1 1 0 2
 1 
G+ =   s−1 s−1
s
0 s+1 
s−1  .
s 1 s+1
0 s−1 s−1 s−1

We get a left coprime polynomial matrix fraction of G+ as

 −1  
1 0 0 1 1 0 2
G+ = (D+ )−1 N + =  0 (s − 1) 0   1 s 0 s + 1 .
0 0 (s − 1) 0 s 1 s+1

G+ has a pole of multiplicity 2 at s = 1 coinciding with a single zero at

s = 1. For this example, since D+ is itself diagonal, we can invariably take
+ + +
Dlm = D+ and Dla = I for any partition so that N + and Dla are always
+ + + +
externally skew prime in C . Obviously, N̄ = N and D̄la = I constitute a
+
dual internally skew prime pair of N + and Dla in C+ . In order to determine
+
the solvability of (P) for this example, we need only to check if Dlm = D+
+ + +
and N̄rm = Nrm are left coprime, where Nrm is a least right block multiple of
N + with respect to a given partition. For the partition, m1 = 2, and m2 = 1,
Example 5.4.3 gives a least right block multiple Nrm (2, 1) as
 
1 0 0
+
Nrm (2, 1) =  1 s − 1 0  .
0 0 1

One easily sees that D+ and Nrm +

(2, 1) are left coprime and (P) is thus
solvable for the partition {2, 1}. On the other hand, if the partition is changed
to m1 = 1, and m2 = 2, then again from Example 5.4.3, a least right block
multiple Nrm (1, 2) corresponding to the present partition is
 
s−1 0 0
+
Nrm (1, 2) =  0 s − 1 0 .
0 0 1

In this case, D+ and Nrm +

(1, 2) are not left coprime and (P) is thus not
solvable for the partition {1, 2}. ♦
178 5. Polynomial Matrix Approach

It has been shown in Section 5.2 that a sufficient condition for solvability
of the diagonal decoupling problem with stability for a square plant is that
the plant has no unstable poles coinciding with zeros and plants generically
satisfy this condition. The same result is established for the general block
decoupling case with arbitrary plants as follows.

Corollary 5.4.2. If a full row rank plant G has no unstable poles coinciding
with zeros, then (P) is solvable for every partition.
+ + + + +
Proof. It follows from Dla D = Dlm that ∆(Dla ) divides ∆(Dlm ) and it
+ + m
further implies, by Lemma 5.4.6, that ∆(Dlm ) divides (∆(D )) . Therefore,
+
under the assumed condition, N + and Dla have no zeros in common and they
are externally skew prime in C and so are in C+ . It follows from Lemmas 5.4.6
+
and 5.4.4 that ∆(Dlm ) divides (∆(D+ ))m and ∆(N̄rm +
) divides (∆(N̄ + ))m =
+ m + +
(∆(N )) , so that square polynomial matrices Dlm and N̄rm have no zeros
in common and they are thus left coprime. Hence, by Theorem 5.4.2, (P) is
solvable. The proof is completed. 
In fact, for the case considered in the above corollary, we can directly
construct a decoupling compensator as follows. Let N + = L+ U be an LU-
factorization and X a polynomial matrix such that U X = I, then take K
as

K = (G− )−1 X adj(L+ )(adj(D+ ))−1 .

Simple calculations yield

GK = a(∆(D+ ))−1 ∆(N + )I, a ∈ R,

so that GK is block decoupled for any partition. Since ∆(D+ ) and ∆(N + )
have been assumed to be coprime in C+ , there are no unstable pole-zero
cancellations in GK. The disadvantage with this decoupling compensator is
that its order may be much higher than necessary. In general, lower order ones
can be obtained from the parameterization of all decoupling compensators,
which will be presented in the next section.
In the extreme case of one block partition, i.e., v = 1, m1 = m, our (P)
defects to nothing than stabilization and the conditions in Theorem 5.4.2
+
should hold automatically for any plant G. Indeed, in this case, Dlm = D+ ,
+ + + +
Dla = I, N̄rm = N̄ = N . The solvability conditions are then reduced to
ones that N + and I are externally skew prime in C+ and D+ and N + are
left coprime, which, of course are satisfied for any G+ .

5.4.4 Characterization of Solutions

Besides the solvability conditions, it is usually required to characterize all

decoupling compensators and achievable loop transfer function matrices so
that free parameters can be used to achieve other performance specifications.
 5.4 A Unified Solution 179

Proposition 5.4.1. If G+ = (D+ )−1 N + satisfies the conditions in Theorem

5.4.2, then all compensators which block-decouple G+ and at the same time
maintain the internal stabilizability are given by
+ −1 +
K = (D̄la ) (N̄ra N̄q1 + ZW )(D̄q1 )−1 , (5.91)
+ +
where D̄la and N̄ra are defined as above, polynomial matrices N̄q1 and D̄q1
are both block diagonal and right coprime, K2 = N̄q1 (D̄q1 )−1 is coprime and
+ −1 + + +
has no pole-zero cancellations with Q1 = (Dlm ) N̄rm , N̄ra and N̄rm satisfy
+ + +
N̄ N̄ra = N̄rm , Z is a minimal polynomial base of the right null space of
N̄ + (Forney,1975) and of size l × (l − m), and W is an arbitrary polynomial
matrix of size (l − m) × m.

Proof. In the proof of necessity of Theorem 5.4.2, one sees from (5.67), (5.69),
(5.70) and (5.82) that every K which block-decouples G+ and maintains
internal stabilizability cab be expressed as
K = Nk Dk−1 = Nk (D̄q1 Dk1 )−1
−1 −1 −1 −1
= Nk Dk1 D̄q1 = D̄k1 N̄k D̄q1
+ −1 −1
= (D̄la ) N̄k D̄q1 . (5.92)
It follows from (5.79) and (5.80) that N̄k satisfies

N̄ + N̄k = N̄rm
+
N̄q1 . (5.93)
+
Since N̄ra N̄q1 is obviously a particular solution of (5.93) for N̄k and ZW is
a general solution of N̄ + N̄k = 0, a general N̄k which satisfies (5.93) is then
given by
+
N̄k = N̄ra N̄q1 + ZW, (5.94)

which is then substituted to (5.92) to yield (5.91). Furthermore, (5.81) can

be expressed as
+ −1 + −1
Q = Dq−1 Nq = (Dlm ) N̄rm N̄q1 D̄q1 = Q 1 K2 . (5.95)
+
As Dq and Nq are coprime and ∆(Dq ) = ∆(Dlm )∆(D̄q1 ) by (5.64) and (5.66),
N̄q1 and D̄q1 are thus coprime and Q1 K2 has no pole-zero cancellations. On
the other hand, following a similar reasoning to the proof of sufficiency of
Theorem 5.4.2 with K1 in (5.85) replaced by K in (5.91), one can easily
show that each K in (5.91) block-decouples G+ and preserves the internal
stabilizability. 
Theorem 5.4.1 says that K is a solution to the block decoupling problem
without unstable pole-zero cancellations for a plant G if and only if so is
K + = G− K for G+ , where G = G+ G− is a stability factorization. But such
K + and the resulting loop maps have been parameterized above. We thus
obtain the following characterization of solutions for G.
180 5. Polynomial Matrix Approach

Theorem 5.4.3. Assume that a full row rank plant G and a partition be
+ + +
given, the notations, G+ , G− , D+ , N + , Dlm , Dla , N̄ + , D̄la +
, and N̄rm are de-
fined as before, and the conditions in Theorem 5.4.2 are satisfied. In addition,
+
let N̄ra +
be a polynomial matrix such that N̄ + N̄ra = N̄rm +
and Z be an l×(l−m)
minimal polynomial base of the right null space of N̄ + , then
(i) all compensators which block-decouple G with internal stabilizability
are given by
+ −1 −1
K = (G− )−1 (D̄la +
) (N̄ra Nk2 + ZW )Dk2 ,

where polynomial matrices Nk2 and Dk2 are both block diagonal with respect
−1
to the partition and right coprime, K2 = Nk2 Dk2 is coprime and has no pole-
+ −1 +
zero cancellations with Q1 = (Dlm ) N̄rm and, W is an arbitrary polynomial
matrix of size (l − m) × m;
(ii) all achievable open-loop transfer function matrices Q = GK under
block decoupling with internal stabilizability can be expressed as

Q = Q 1 K2 ,

where Q1 and K2 are the same as in (i).

From the point of view of closed-loop system design, all remains to do is
construct a block diagonal stabilizer K2 for the decoupled plant Q1 . It be-
comes a sequence of reduced size design problems. In addition, Theorem 5.4.3
exhibits useful information about the zero and pole properties of a decoupled
system. Because there are no unstable pole-zero cancellations between Q1
and K2 , then Q1 will thus entirely remain in each decoupled loop for a given
plant G no matter what K2 is chosen. It follows from Lemmas 5.4.6 and 5.4.4
+
that ∆(D+ ) divides ∆(Dlm ) and ∆(N + ) divides ∆(N̄rm
+
), but Q1 introduces
neither zeros nor poles different from those of the plant. What may happen
is that multiplicities of zeros or poles of Q1 are greater than those of G. It
is interesting to find out a class of plants for which the multiplicities are not
increased, that is,
+
∆(Dlm ) = ∆(D+ ) (5.96)

and
+
∆(N̄rm ) = ∆(N + ) (5.97)

are both satisfied.

Theorem 5.4.4. Let a full row rank plant G and a partition be given, then
the following statements are equivalent:
(i) (P) is solvable, and both (5.96) and (5.97) hold true.
(ii) There exists an irreducible polynomial matrix U such that
+ −1 +
(Dlm ) (N̄rm U ) is a left coprime polynomial matrix fraction of G+ .
 5.4 A Unified Solution 181

(iii) There is a stable and minimum phase compensator K such that GK

is block decoupled and ∆(N + ) = ∆+ (N + Nk ), where Nk Dk−1 is a coprime
polynomial matrix fraction of G− K.

Proof. (i) ⇒ (ii): If (P) is solvable and both (5.96) and (5.97) hold true,
+ +
then ∆(Dla ) = 1 and Dla is unimodular. The C+ -skew primeness equa-
+ + + + +
tion, Dla N = N̄ D̄la , can be thus rewritten as N + = P D̄la , where
+ −1 + +
P = (Dla ) N̄ is a polynomial matrix. Note that N is antistable, and
+ + +
D̄la is stable since, by Lemma 5.4.4, ∆+ (D̄la ) = ∆(Dla ) = 1. It then
+ −1
follows from Lemma 5.4.2 that U1 (D̄la ) = P1 is a polynomial matrix,
where N + = L+ U1 is an LU-factorization. Consider now the equality,
+ −1 + + −1 +
(D+ )−1 N + (D̄la ) N̄ra = (Dlm ) N̄rm . It can be rewritten as
+ −1 +
+
(D+ )−1 L+ P1 N̄ra = (Dlm ) N̄rm . (5.98)
+
Using (5.96) and (5.97) with ∆(L+ ) = ∆(N + ) and the coprimeness of Dlm
+ +
and N̄rm , one concludes that U2 = P1 N̄ra is a unimodular polynomial ma-
trix. Take U as U = U2−1 U1 , then U is an irreducible polynomial matrix.
Postmultiplying (5.98) by U yields
+ −1 +
(Dlm ) N̄rm U = (D+ )−1 L+ U1 = (D+ )−1 N + = G+ . (5.99)
+ −1
In view of (5.96) and (5.99) with D+ and N + left coprime, (Dlm +
) (N̄rm U)
is thus left coprime, too.
(ii) ⇒ (iii): Let X be a polynomial matrix such that U X = I, then X
has neither zeros nor poles and K = (G− )−1 X is stable and minimum phase.
The resulting loop transfer matrix becomes
+ −1 + + −1 +
Q = GK = (Dlm ) N̄rm U G− (G− )−1 X = (Dlm ) N̄rm ,

which is block decoupled. In this case, N + = N̄rm

+
U and Nk = X. Hence,
∆(N + ) = ∆(N̄rm
+
) = ∆+ (N + Nk ).
(iii) ⇒ (i): Let N + = L+ U1 be an LU-factorization, then (iii) implies
that the square matrix U1 Nk Dk−1 is stable and minimum phase and

[(D+ )−1 L+ ][U1 Nk Dk−1 ] = GK (5.100)

has no unstable pole-zero cancellations. Such a K block-decouples G while

maintaining stabilizability and the resulting loop transfer matrix is thus char-
acterized in Theorem 5.4.3 so that
+
∆(N̄rm )∆+ (Nk2 ) = ∆+ (Nq ), (5.101)

where Dq−1 Nq is a coprime polynomial matrix fraction of Q = GK. It follows

from (5.100) that ∆+ (Nq ) divides ∆+ (N + Nk ) = ∆(N + ), where the equal-
ity is the given condition. This, together with (5.101), further implies that
182 5. Polynomial Matrix Approach

+
∆(N̄rm ) divides ∆(N + ). But by Lemmas 5.4.4 and 5.4.6, ∆(N + ) divides
+
∆(N̄rm ), too. Hence, (5.97) holds true. (5.96) can be proven in a similar way
by noting from (5.100) that ∆+ (Dq ) divides ∆(D+ )∆+ (Dk ) = ∆(D+ ) as K
is stable. The proof is completed. 

Example 5.4.5. Consider again the plant in Example 5.4.4. For the partition
{2, 1}, Example 5.4.4 has shown that (P) is solvable. In addition, one easily
+ +
sees that ∆(N̄rm ) = (s − 1) = ∆(N + ) and ∆(Dlm ) = ∆(D+ ) = (s − 1)2 .
Therefore, the conditions in (i) of Theorem 5.4.4 are satisfied. G+ can be
expressed as
+ −1 +
G+ = (Dlm ) N̄rm U =
 −1   
1 0 0 1 0 0 1 1 0 2
 0 (s − 1) 0  1 s − 1 0 0 1 0 1 ,
0 0 (s − 1) 0 0 1 0 s 1 s+1

which is in the form given by (ii) of Theorem 5.4.4. Take K as

 
1 −1 0
 0 1 0 
K = (G− )−1   0 −s
,
1 
0 0 0

then K is stable and minimum phase and

 
1 0 0
GK = (Dlm ) N̄rm =  s−1
+ −1 + 1
1 0 ,
1
0 0 s−1

which is block decoupled with respect to the partition {2, 1}. One also sees
that N + = Nrm +
U and N + Nk = Nrm +
so that ∆(N + ) = ∆+ (N + Nk ) =
+
∆(Nrm ). Thus, (iii) of Theorem 5.4.4 also holds, indeed. ♦

It should be pointed out that (i) or (ii) of Theorem 5.4.4 implies that
there is a stable and minimum phase compensator K such that GK is block
decoupled and has no unstable pole-zero cancellations, but the converse is
not true in general. For instance, consider the plant:
 s 1

s−1 s−1 0
G=
0 1 0
 −1  
s−1 0 s 1 0
=
0 1 0 1 0

= (D+ )−1 N + ,
for the partition {1, 1}. Take K as
 5.4 A Unified Solution 183

 
1 −1  −1
  1 0
K= 0 s = Nk Dk−1 ,
0 s+1
0 1

which is stable and minimum phase. One can verify that GK is decoupled
and has no unstable pole-zero cancellations. Simple calculations show that
∆(N + ) = s but ∆+ (N + Nk ) = s2 so that ∆(N + ) 6= ∆+ (N + Nk ) for this K.
Are there any other decoupling K which makes the equality

∆(N + ) = ∆+ (N + Nk ) (5.102)

hold? The answer is No! To see this, one notes from Corollary 5.4.2 that (P) is
solvable for the given plant and the partition {1, 1}, and (5.101) thus applies
+ +
with N̄rm = Nrm = diag{s, s}. It follows that deg + (N + Nk ) ≥ deg + (Nq ) ≥
deg (Nrm ) = 2 > 1 = deg(N + ). Therefore, among all the compensators K
+ +

such that GK is block decoupled without unstable pole-zero cancellations,

there is no one satisfying (5.102). One concludes from this example that
the condition that there is a stable and minimum phase compensator which
block-decouples the plant with internal stabilizability is only necessary but
not sufficient for either (i) or (ii) of Theorem 5.4.4 to hold. In general, the
additional condition (5.102) is needed to make it equivalent to (i) or (ii) of
that theorem. However, if the plant is square, (5.102) will be always satisfied
provided K is of minimum phase.

Corollary 5.4.3. If the plant G is square, (iii) of Theorem 5.4.4 can be

replaced by the following: there is a stable and minimum phase compensator
K such that GK is block decoupled.

The decoupling compensators constructed so far usually result in sta-

ble pole-zero cancellations, G− (G− )−1 , so that arbitrary pole assignment is
not achievable. However, by imposing the stronger constraint, no pole-zero
cancellations in GK, instead of only RHP ones, we can directly obtain the
following result and the proof is omitted because of its similarity to those of
Theorems 5.4.2 and 5.4.3.

Theorem 5.4.5. Given a full row rank plant G ∈ Rm×l p (s) and a partition
{mi }, let D−1 N = G be a left coprime polynomial matrix fraction and Dlm
a least left block multiple of D with respect to the partition such that Dlm =
Dla D for some polynomial matrix Dla . If N and Dla are externally skew
prime, construct internally skew prime polynomial matrices Ñ and D˜la such
that Dla N = Ñ D˜la . And let Ñrm be a least right block multiple of Ñ with
respect to the partition such that Ñ N˜ra = N˜rm for some polynomial matrix
N˜ra and Z a l × (l − m) minimal polynomial base of the right null space of
Ñ , then the following hold true:
(i) there is a proper compensator K in a unity output feedback system
such that GK is nonsingular and block diagonal with respect to the partition
184 5. Polynomial Matrix Approach

and arbitrary closed-loop pole assignment is achievable if and only if N and

Dla are externally skew prime and Dlm and Ñrm are left coprime;
(ii) under the conditions of (i), all compensators which block-decouple G
and preserve arbitrary pole assignability are given by
−1 −1
K = D̃la (Ñra Nk2 + ZW )Dk2 ,

where polynomial matrices Nk2 and Dk2 are both block diagonal with respect
to the partition and right coprime, Q1 K2 has no pole-zero cancellations with
−1 −1
Q1 = Dlm Ñrm and K2 = Nk2 Dk2 , and W is an arbitrary polynomial matrix
of size (l − m) × m; And
(iii) all achievable open-loop maps under block decoupling with arbitrary
pole assignability can be expressed as

Q = Q 1 K2 ,

where Q1 and K2 are the same as in (ii).

In this section, the block decoupling problem with internal stability for
unity output feedback systems has been considered in a general setting and
completely solved. Necessary and sufficient conditions for its solvability are
given and are only concerned with the greatest C + -divisor of a given plant.
The characterization of decoupled systems is presented and it shows that a
general compensator which solves the problem consists mainly of two parts,
one being the decoupling part which makes the plant block decoupled with-
out unstable pole-zero cancellations, the other the stabilizing part which is
block diagonal and makes the block decoupled plant internally stable and
the whole compensator proper. The block decoupled plant (the plant plus
the decoupling part) need not introduce different zeros or poles from those
of the plant itself and what may happen is that zero or pole multiplicities of
the former are greater than those of the latter. Two new concepts, least block
multiples of a full rank polynomial matrix and skew prime polynomial matri-
ces in C+ , are introduced and they have played important roles, respectively,
in the problem solvability and compensator parameterization.

5.5 Notes and References

The existence condition for decoupling by unity output feedback compen-

sation is that the plant has full row rank and it is very simple and well
known. But the problem of decoupling in such a configuration while main-
taining internal stability of the decoupled system appears to be much more
difficult. Early papers on the subject (Safonov and Chen, 1982; Vardu-
lakis, 1987; Peng, 1990; Desoer, 1990; Lin and Hsieh, 1991) consider the
special case that the set of RHP zeros of the system is disjoint from the set of
 5.5 Notes and References 185

poles. They show that the diagonal decoupling problem is solvable under this
assumption and provide a parameterization of controllers solving the prob-
lem This assumption is relaxed and the diagonal decoupling problem with
stability is solved by Linnemann and Maier (1990) for 2 × 2 plants, and by
Wang (1992) for square plants, along with a controller parameterization. For
the same problem, alternative necessary and sufficient conditions are also de-
rived by Lin (1997) based on transfer matrices and residues, and in Wang
and Yang (2002) using the minimal C + -decoupler of the plant. The results
of Wang (1992) are generalized to the block decoupling problem for possibly
non-square plants by Linnemann and Wang (1993) with additional discus-
sion on the performance limitations in the decoupled system compared with
the performance which is achievable by not necessarily decoupling controllers.
A unified and independent solution for the general case is also presented
in Wang (1993). The main tools used in this chapter are skew primeness
(Wolovich, 1978), strict adjoints (Hammer and Khargonekar, 1984), stability
factorizations (Wang, 1992), and their extensions to non-square plants and
block diagonalization.
6. Transfer Matrix Approach

In the preceding two chapters, the plants which we have considered assume
no time delay and for such a kind of plants the complete solution has been
developed in both state-space and polynomial matrix settings. However, time
delay often exists in industrial processes and other systems, some of which will
be highlighted in Section 1. We need an entirely new approach to decoupling
problem for plants with time delays. Time delay imposes a serious difficulty
in theoretical development and thus the theory of time delay systems is much
less matured than delay-free systems. In the present and next chapters, we
will address decoupling problem for plants with time delay and our emphasis
is to develop decoupling methodologies with necessary theoretical supports
as well as the controller design details for possible practical applications.
This chapter will consider conventional unity output feedback configuration
whereas the next chapter will exclusively deal with time delay compensation
schemes.
In this chapter, several typical plants with time delay will be briefly de-
scribed and their transfer matrices with time delay given in Section 1 as the
motivation of subsequent discussions on delay systems. The new decoupling
equations are derived in a transfer function matrix setting for delay plants
(Section 2), and achievable performance after decoupling is analyzed (Section
3) where characterization of the unavoidable time delays and non-minimum
phase zeros that are inherent in a feedback loop is given. In Section 4, a
design method is presented. The objective is to design a controller of lowest
complexity which can achieve fastest loops with acceptable overshoot and
minimum loop interactions. Stability and robustness are analyzed in Section
5. Section 6 shows why multivariable PID control often fails when the number
of inputs/outputs is getting bigger. More simulation is given in Section 7 and
the conclusion is drawn in Section 8.

6.1 Plants with Time Delay

In the literature of process control, there are many 2 × 2 process models.
But usually they are subsystems of actual processes. A typical process unit
in industry has number of inputs/output well beyond 2, some can reach
more than 10. And it usually has some time delay. A number of industrial
Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 187-235, 2003.
Springer-Verlag Berlin Heidelberg 2003
188 6. Transfer Matrix Approach

processes are introduced in this section to show the presence of time delay
and will be used in simulation later to substantiate the effectiveness of the
proposed decoupling method for high performance control.

Example 6.1.1. Tyreus (1982) studied a sidestream column separating a

ternary mixture, where a feed containing 10 percent benzene, 45 percent
toluene and 45 percent o-xylene is separated in a single column into three
product streams: a benzene distillate with 5 percent toluene impurity, a liquid
toluene sidestream with 5 percent benzene and 6 percent xylene impurities,
and a xylene bottom with 5 percent toluene impurity. The transfer function
matrix for the column is
 −0.71s −60s −2.24s

1.986e −5.24e −5.984e
 66.7s+1 400s+1 14.29s+1

 −0.59s
0.33e−0.68s −2.38e−0.42s 
G(s) =  −0.0204e 2 (2.38s+1)2 (1.43s+1)2 , (6.1)
 (7.14s+1) 
−7.75s −3.79s −1.59s
−0.374e 11.3e 9.811e
22.22s+1 (21.74s+1)2 11.36s+1

where the controlled and manipulated variables are y1 (toluene impurity in

the distillate); y2 (benzene impurity in the sidestream); y3 (toluene impurity
in the bottom); u1 (reflux ratio); u2 (sidestream flow rate); u3 (reboil duty).
♦

Example 6.1.2. Doukas and Luyben (1978) studied the dynamics of a dis-
tillation column producing a liquid sidestream product, with the objective
of maintaining four composition specifications on the three product streams.
The Tyreus process in (6.1) is a 3 × 3 subsystem of it. The transfer function
matrix for the 4 × 4 model is
 −11.3e−3.79s 0.374e−7.75s −9.811e−1.59s −2.37e−27.33s 
(21.74s+1)2 22.2s+1 11.36s+1 33.3s+1
 
 5.24e−60s −1.986e−0.71s 5.984e−2.24s 0.422e−8.72s 
 400s+1 66.67s+1 14.29s+1 (250s+1)2 ; 
G(s) = 
 −0.33e−0.68s
,
 (6.2)
0.0204e−0.59s 2.38e−0.42s
 (2.38s+1)2 0.513e−s 
 (7.14s+1)2 (1.43s+1)2 
4.48e−0.52s −0.176e−0.48s −11.67e−1.91s
11.11s+1 (6.90s+1)2 12.19s+1 15.54e−s

where the controlled and manipulated variables are y1 (toluene impurity in

the bottom); y2 (toluene impurity in the distillate); y3 (benzene impurity in
the sidestream); y4 (xylene impurity in the sidestream); u1 (sidestream flow
rate); u2 (reflux ratio); u3 (reboil duty); and u4 (side draw location). ♦

Example 6.1.3. Alatiqi and Luyben (1986) presented the results of a quanti-
tative study of the dynamics of two alternative distillation systems for sep-
arating ternary mixtures that contain small amounts (less than 20%) of the
intermediate component in the feed. The process transfer function matrix of
the complex sidestream column/stripper distillation process is given by
 6.1 Plants with Time Delay 189

 4.09e−1.3s −6.36e−0.2s −0.25e−0.4s −0.49e−5s


(33s+1)(8.3s+1) (31.6s+1)(20s+1) 21s+1 22s+1
 
 −4.17e−4s 6.93e−1.01s −0.05e−5s 1.53e−2.8s 
 45s+1 44.6s+1 34.5s+1 48s+1 
G(s) = 

,

−1.73e−17s 5.11e−11s 4.61e−1.02s −5.48e−0.5s
 
 13s+1 13.3s+1 18.5s+1 15s+1 
−11.18e−2.6s 14.04e−0.02s −0.1e−0.05s 4.49e−0.6s
(43s+1)(6.5s+1) (45s+1)(10s+1) (31.6s+1)(5s+1) (48s+1)(6.3s+1)
(6.3)

where the four controlled variables are y1 (mole fraction of the first component
in the distillate); y2 (mole fraction of the third component in the bottom); y3
(mole fraction of the second component in the sidestream); y4 (temperature
difference); and the four manipulated variables are u1 (reflux flow rate); u2
(main column reboiler heat-transfer rate); u3 (stripper reboiler heat-transfer
rate); and u4 (liquid draw rate from main column to stripper). Although the
complex sidestream column/stripper configuration is more energy efficient
than the conventional two-column sequential ”light-out-first” configuration,
it presents a challenging 4 × 4 multivariable control problem. ♦
Example 6.1.4. Ammonia is commercially produced by synthesis of the raw
material gas (Hydrogen, H2 , Nitrogen, N2 ) via a reversible reaction catalyt-
ically with great energy consumption for making raw gases. The recycled
un-reacted raw gas makes the whole ammonia process more difficult to oper-
ate and control. In general, ammonia process consists of two big sessions. One
is reformer session to produce reactant gas (H 2 , N2 ). Another is conversion
session to synthesize ammonia. Many disturbances and frequent change of
feed gas rate affect a modern ammonia plant. Moreover, these disturbances
always propagate throughout the entire plant. It has been common require-
ments to improve operating stability, increase material and energy efficiency
and prevent disturbances from echoing throughout the plant. A transfer func-
tion matrix of the reformer session was identified as
 −4s −s −4s

−0.562e −0.01e 0.378e
 8s+1 3s+1 8s+1

 −19s
−0.0159e−21s  ,
G(s) =  0.0135e
5.6s+1 0 10s+1  (6.4)
 
−10s −4s −7.5s
0.002e −0.00125e 0.12e
7s+1 3s+1 10s+1

where the controlled variables and manipulated variables are y1 (the primary
reformer coil outlet temperature); y2 (the secondary reformer methane leak-
age) and y3 (the H2 /N2 ratio); u1 (the process natural gas flow rate); u2 (the
process air flow rate); and u3 (the fuel natural gas flow rate). ♦
Example 6.1.5. The purpose of a depropanizer column is to separate propane
from the feed which comes from deethanizer column. The operation target of
this unit is to keep the top butane concentration low than 5% and bottom
propane concentration low than 5%. The main controlled variables includes
y1 (top butane concentration), y2 (bottom propane concentration) and y3
190 6. Transfer Matrix Approach

(column DP flooding). The main manipulated variables are u1 (column top

reflux flow), u2 (Column bottom steam flow), u3 (column overhead pressure).
The 3 × 3 transfer function matrix model was found as
 −27.5s −53.5s −56s

−0.26978e 1.978e 0.07724e
97.5s+1 118.5s+1 96s+1
 
 0.4881e−117s −5.26e−26.5s 0.19996e−35s  .
G(s) =  56s+1 58.5s+1 51s+1  ♦ (6.5)
 
0.6e−16.5s 5.5e−15.5s −0.5e−17s
40.5s+1 19.5s+1 18s+1

We have so far highlighted several typical complex industrial processes

that appear in the literature or that we encountered during our industrial
control practice in developing and implementing a multivariable controller.
One should note that the complexity of the above processes do not come
from the dynamics of the individual channels of the processes as clearly they
are mostly nothing more than first or second order plus time delay form, but
is from the multivariable interactions existing between the various variables.
The more interactive loops are involved, the more difficult it is to compensate
for these interactions. The difficulty of interaction handling and the demand
for better decoupling is evidenced by the recent survey (Kong, 1995) from the
leading international control companies such as Fish-Rosemount, Yokogawa
and Foxboro, all of whom ranked poor decoupling as the principal common
control problem in industry. Poor decoupling is universal and in most cases,
the closure of a multivariable feedback loop brings more interactions to the
controlled variable than the open-loop operation, thus leaving the benefit of
the close-loop control only as mere preventing the controlled variable from
flowing away. Therefore, we are well motivated to consider decoupling prob-
lem for plants with time delay.

6.2 Decoupling

Consider the conventional unity feedback control system in Figure 6.1, where
G represents the transfer matrix of the plant and K the multivariable con-
troller. Here, G is assumed to be square and nonsingular:

Fig. 6.1. Unity Output Feedback System

 6.2 Decoupling 191

 
g11 (s) · · · g1m (s)
 ..  ,
G(s) =  ... ..
. .  (6.6)
gm1 (s) · · · gmm (s)

where

gij (s) = gij0 (s)e−Lij s ,

and gij0 (s) are strictly proper, stable scalar rational functions and Lij are
non-negative constants, representing time delays. The controller K(s) is an
m × m full-cross coupled multivariable transfer matrix:
 
k11 (s) · · · k1m (s)
 ..  ,
K(s) =  ... ..
. .  (6.7)
km1 (s) · · · kmm (s)

where

kij (s) = kij0 (s)e−θij s ,

θij are non-negative constants, and kij0 (s) are scalar proper rational func-
tions. kij0 (s) should be stable except a pole at s = 0, and for simplicity, it is
called stable in this chapter (referred to the controller only). This pole is de-
sired to have integral control. Assume that the plant has no zero at the origin.
Then, the integral control will lead to zero steady state errors in response to
step inputs provided that the closed-loop is stable. It is also assumed that a
proper input-output pairing has been achieved in G(s) such that none of the
m principal minors of G(s) (the i th principle minor is the determinant of a
(m − 1) × (m − 1) matrix obtained by deleting the i th row and i th column
of G(s)) are zero.
Our task here is to find a K(s) such that the closed loop transfer matrix
from the reference vector r to the output vector y:
−1
H(s) = G(s)K(s) [I + G(s)K(s)] , (6.8)
is decoupled, that is, H(s) is diagonal and nonsingular. One sees that
H −1 = [I + GK] (GK)−1
= (GK)−1 + I, (6.9)
and the closed-loop H(s) is decoupled if and only if the open loop G(s)K(s)
is decoupled, i.e.,
 
g1
 g2 
 
GK =  .  [k1 k2 · · · km ] = diag{qii }, i = 1, 2, . . . , m. (6.10)
 .. 
gm
192 6. Transfer Matrix Approach

For each column of GK, we have

 
g1
 g2   T
 
 ..  ki = 0 · · · 0 qii 0 · · · 0 , i = 1, 2, . . . , m, (6.11)
 . 
gm
which is equivalent to
 
g1
 .. 
 . 
 
 gi−1 
 
 gi+1  ki = 0, i = 1, 2, . . . , m, (6.12)
 
 . 
 .. 
gm
and
gi ki = qii 6= 0, i = 1, 2, · · · , m, (6.13)
For any i, we can solve (6.12) to obtain k1,i , · · · , ki−1,i , ki+1,i , · · · , km,i , in
terms of kii as
   
k1i ψ1i
 ..   .. 
 .   . 
   
ki−1,i  ψi−1,i 
 = 
ki+1,i  ψi+1,i  kii , ∀ i ∈ m, (6.14)
   
 .   . 
 ..   .. 
kmi ψmi
where m = {1, 2, · · · , m} and
   −1  
ψ1i g11 · · · g1,i−1 g1,i+1 · · · g1m g1i
 ..   .. .. .. .. .. ..   .. 
 .   . . . . . .   
     . 
ψi−1,i  gi−1,1 · · · gi−1,i−1 gi−1,i+1  
· · · gi−1,m  gi−1,i 
   .
ψi+1,i  , − gi+1,1 · · · gi+1,i−1 gi+1,i+1 · · · gi+1,m   
    gi+1,i 
 .   . .. .. .. .. ..   . 
 ..   .. . . . . .   .. 
ψmi gm1 · · · gm,i−1 gm,i+1 · · · gmm gmi
(6.15)
The i th diagonal element of GK is g̃ii kii , where
m
X
g̃ii = gii + gik ψki . (6.16)
k=1
k6=i
 6.2 Decoupling 193

One notes that for a given i, the resulting diagonal element of GK is indepen-
dent of the controller off-diagonal elements but contains only the controller
diagonal element kii .
To further simplify the above result, let Gij be the cofactor corresponding
to gij in G. It follows from linear algebra (Noble, 1969) that the inverse of G
 ji 
can be given as G−1 = adj G
|G| , where adj G = G . This also means
   11 
g11 · · · g1m G · · · Gm1
 .. . . ..   .. .. . 
 . . .  . . ..  = |G|Im . (6.17)
gm1 · · · gmm G1m · · · Gmm

Equation (6.17) can be equivalently put into the following two relations:
   i1   
g11 · · · g1,i−1 g1,i+1 · · · g1m G g1i
 .. .. .. .. .. ..   ..   .. 
 . . . . . .   .   . 
   i,i−1   
gi−1,1 · · · gi−1,i−1 gi−1,i+1 · · · gi−1,m  G   
   i,i+1  = − gi−1,i  Gii , ∀ i ∈ m,
gi+1,1 · · · gi+1,i−1 gi+1,i+1 · · · gi+1,m  G  gi+1,i 
    
 . .. .. .. .. ..   ..   . 
 .. . . . . .   .   .
. 
gm1 · · · gm,i−1 gm,i+1 · · · gmm Gim gmi
(6.18)

and
m
X
gik Gik = |G|, ∀i ∈ m.
k=1

Substituting (6.18) to (6.15) yields

   −1    i1 
ψ1i g11 ··· g1,i−1 g1,i+1 · · · g1m g1i G
 .   . .. .. .. .. ..   ..   .. 
 ..   .. . . . . .     . 
     .   i,i−1 
ψi−1,i   · · · gi−1,m  gi−1,i  1 G 
  = − gi−1,1 ··· gi−1,i−1 gi−1,i+1   =  i,i+1  ,
ψi+1,i  gi+1,1 ··· gi+1,i−1 gi+1,i+1 · · · gi+1,m  gi+1,i  G G
ii  
      
 .   . .. .. .. .. ..   ..   . 
 ..   .. . . . . .   .   .
. 
ψmi gm1 ··· gm,i−1 gm,i+1 · · · gmm gmi Gim

or
Gij
ψji = , ∀ i, j ∈ m, j 6= i. (6.19)
Gii
By (6.19), (6.14) and (6.16) become respectively

Gij
kji = kii , ∀ i, j ∈ m, j 6= i, (6.20)
Gii
194 6. Transfer Matrix Approach

and
m
X m
Gik 1 X |G|
g̃ii = gii + gik ii
= ii
gik Gik = ii , ∀ i ∈ m. (6.21)
G G G
k=1 k=1
k6=i

Thus, the decoupled open-loop transfer function matrix is given by

 
|G|
GK = diag {g̃ii kii } = diag kii , i = 1, 2, · · · , m . (6.22)
Gii
ii
Before proceeding further, it is interesting to note that G |G| is the (i, i )-th
element of G−1 . And it follows from Bristol (1966) that the (i,i )-th element
ii
of the relative gain array for G is given by λii = gii G|G| . Hence, g̃ii is linked
to gii by g̃ii = λ−1
ii gii .
To demonstrate how to find g̃ii and ψji , take the following as an example:
 1 e−2s −1 e−6s 
 s + 2 s +2 
G(s) =  2 . (6.23)
s − 0.5 e−3s (s − 0.5) e−8s
(s + 2)2 2(s + 2)3

Simple calculations give

2(s − 0.5)(s + 2)e−9s + (s − 0.5)2 e−10s

|G| = ,
2(s + 2)4

(s − 0.5)2 e−8s e−6s

G11 = , G21 = ,
2(s + 2)3 s+2

e−2s (s − 0.5)e−3s
G22 = , G12 = − .
s+2 (s + 2)2

It follows from (6.21) that the decoupled loops have their equivalent processes
as
2(s + 2)e−s + (s − 0.5)e−2s
g̃11 = ,
(s − 0.5)(s + 2)

2(s − 0.5)(s + 2)e−7s + (s − 0.5)2 e−8s

g̃22 = ,
2(s + 2)3

respectively. By (6.19), ψ21 and ψ12 are found as

2(s + 2) 5s
ψ21 = − e , ψ12 = e−4s .
s − 0.5
 6.3 Achievable Performance 195

6.3 Achievable Performance

A natural question to ask after decoupling the plant is what the achievable
performance of the decoupled system is. It is well known that the performance
limitations are imposed by the non-minimum phase part of the system. This
requires us to derive the characterization of the unavoidable time delays and
non-minimum phase zeros that are inherent in any feedback loop of the de-
coupled plants. Furthermore, to be applicable in practice, we also have to
find a realizable controller, preferentially of lowest complexity, which can
yield the best achievable performance. Note from the preceding section that
the exact decoupler is likely to be very complex and thus not practical for
industrial implementation. Rather, one would prefer to have a controller of
normal type in the format of rational function plus possible time delay to get
near-decoupling performance. This near-decoupling would usually benefit in-
dividual loop performance as there is a trade-off between loop-decoupling and
loop performance in most cases. These considerations lead us to the following
three-stage decoupling design approach.
(i) Derive the best achievable objective function for each decoupled closed-
loop transfer function hri (s) based on the process dynamic character-
istics. Let the performance specifications of the decoupled closed-loop
system are formulated as m objective closed-loop transfer functions hri
from the set-point ri to output yi , ∀ i ∈ m. Then, the ith open-loop
transfer function qrii corresponding to the objective closed-loop transfer
function hri is given by
hri
qrii = . (6.24)
1 − hri
(ii) Match the actual loop g̃ii kii in (6.22) to the objective loop:

g̃ii (s)kii (s) = qrii (s),

to find the ideal diagonal elements of the controller as

IDEAL −1 Gii (s)

kii (s) = g̃ii (s)qrii (s) = qrii (s), (6.25)
|G(s)|
and the corresponding ideal decoupling off-diagonal elements kji (s) of
the controller as

IDEAL Gij (s) IDEAL Gij (s)

kji (s) = kii = qrii (s). (6.26)
Gii (s) |G(s)|
The ideal controller is generally highly complicated and/or difficult to
realize.
(iii) Approximate the ideal controller kij (s) element-wisely with a simplest
rational function plus possible delay with a pre-specified accuracy.
196 6. Transfer Matrix Approach

In the present section, we will only consider analytical characteristics of

the achievable objective loop transfer function whereas the rest of the above
approach will be addressed in the next section, after which there will be
stability and robustness analysis and simulation results.
The achievable control performance for a given process depends on its
characteristics. Thus, the objective transfer functions cannot be chosen arbi-
trarily. Assume that the ith loop control specifications are expressed in terms
of a desired closed-loop transfer function:
2 −Li s
ωni e Y  z − s ni (z)
hri (s) = 2 2 )( 1 s + 1)νi
, i ∈ m.
(s + 2ωni ξi s + ωni Ni ωni
z+s
z∈Zi
(6.27)

Notice that time delays and non-minimum phase zeros are inherent charac-
teristic of a process and can not be altered by any feedback control. They
Q  ni (z)
are reflected in (6.27) respectively by e−Li s and z∈Zi z−s z+s . The fac-
tor of the standard 2nd-order transfer function represents usual performance
requirement such as overshoot and settling time for the minimum phase and
assignable part of the loop. The term ( Ni1ωni s + 1)νi is to provide necessary
high frequency roll-off rate required by decoupling and properness of the
controller.
Firstly, consider the time delay Li of the ith loop. As time delay imposes
performance limitation in a control loop, to achieve best performance and
to make the closed-loop response fastest, the time delay Li in the ith loop
in (6.27) should be chosen as the smallest possible value. If only (6.22) is
concerned, it seems that Li can be chosen as the time delay in g̃ii as kii cannot
contain pure prediction for realizability. However, decoupling may require this
delay to be increased. To see this, assume that the system interaction from
ith input to jth output, gji , has smaller time delay than that of the transfer
function from the jth input to the jth output. Then, complete elimination
of the interaction, gji , by a realizable kji is impossible, unless the ith loop
delay is increased by a certain amount.
It is thus interesting to know when decoupling requires further delay in
a certain loop and by what amount. To address this problem, consider the
process G(s) given by (6.6). A general expression for non-zero Gij and |G|
will be
M
X
φ(s) = φl (s)e−αl s , αl > 0, (6.28)
l=1

where φl (s), l = 1, · · · , M , are non-zero scalar rational transfer functions.

Define the time delay for φ(s) in (6.28) as

τ (φ(s)) = min (αl ).

l=1,··· ,M
 6.3 Achievable Performance 197

It is easy to verify τ (φ1 φ2 ) = τ (φ1 )+τ (φ2 ) for any non-zero φ1 (s) and φ2 (s).
And it is obvious that for any realizable and non-zero φ(s), τ (φ(s)) can not
be negative. It follows from (6.26) that
IDEAL
|G|kji = Gij qrii , ∀ i, j ∈ m. (6.29)

Thus, a realizable controller requires

τ (Gji ) + τ (qrii ) ≥ τ (|G|), ∀ i ∈ m, j ∈ Ji ,

where Ji , {j ∈ m | Gij 6= 0}, or

τ (qrii ) ≥ τ (|G|) − τi , ∀ i ∈ m, (6.30)

where

τi , min τ (Gij ). (6.31)

j∈Ji

Therefore, the time delay Li in the ith objective closed-loop transfer function
is chosen as the minimum among all meeting (6.30):

Li = τ (|G|) − τi . (6.32)

We have just seen how the various time delays in Gij and |G| limit the
control system performance and affect the choice of the objective loop transfer
function. In addition to time delays, zeros of Gij and |G| which lie in the
closed right half plane (RHP) may also limit the control system performance.
To address this problem, let Z+ 1
|G| be the set of RHP zeros of |G|. For a stable
and proper kji , |kji (z)| is finite for all Re(z) > 0. Thus, the left side of (6.29)
will evaluate to 0 for each z ∈ Z+ ij
|G| , and hence, the right side of (6.29), G qrii
+
must also interpolate to 0 for each z ∈ Z|G| . This implies that qrii needs to
have a RHP zero at z if z is itself not a RHP zero of Gij . The situation may
become more complicated when RHP zeros of |G| coincide with those of Gij
and when multiplicity is considered.
In order to investigate what RHP zero must be included in qrii , for a
non-zero transfer function φ(s), let η z (φ) be a non-negative integer ν such
that lims→z φ(s)/(s − z)ν exists and is non-zero. Thus, φ(s) has η z (φ) zeros
at s = z if η z (φ) > 0, or no zeros if η z (φ) = 0. It is easy to verify that
η z (φ1 φ2 ) = η z (φ1 ) + η z (φ2 ) for any non-zero transfer functions φ1 (s) and
IDEAL
φ2 (s). It then follows from (6.29) that a stable and proper kji (s) requires

η z (Gij qrii ) ≥ η z (|G|), ∀ i ∈ m, j ∈ Ji , z ∈ Z+

|G| ,

1
It is assumed in this chapter that |G| has no zero at the origin so as to enable
the inclusion of integral control in K(s) for robust tracking and regulation with
respect to step inputs.
198 6. Transfer Matrix Approach

or

η z (qrii ) ≥ η z (|G|) − ηi (z), ∀ i ∈ m, z ∈ Z+

|G| , (6.33)

where

ηi (z) , min η z (Gij ). (6.34)

j∈Ji

Equation (6.33) implies that the ith decoupled open loop transfer func-
tion qrii must contain the RHP zeros at each Z+ |G| with multiplicity of
η z (|G|) − ηi (z) if η z (|G|) − ηi (z) > 0. Note that a feedback connection of
a loop transfer function alter neither the existence nor the location of a zero.
Thus the ith decoupled closed-loop transfer function hri should also contain
these RHP zeros with exactly the same multiplicity. Note also from (6.29)
and the assumed stability of G that qrii need not include any other RHP zero
except those of |G|; and for best performance of the closed-loop system it is
undesirable to include any other RHP zeros in qrii (s) more than necessary.
Therefore, Zi and ni (z) in (6.27) are chosen respectively as

Zi = Z+
|G| , (6.35)

and

ni (z) = η z (|G|) − ηi (z). (6.36)

We now consider the term ( Ni1ωni s + 1)νi , which provides necessary high
frequency roll-off rate required by decoupling and properness of the controller.
Ni is usually chosen as 10 ∼ 20. To determine the integer νi , consider a
non-zero transfer function φ(s), let ν(φ) be the smallest integer ν such that
limω→∞ |φ(jω)/(jω)(1+ν) | = 0. A transfer function φ is said to be proper if
ν(φ) ≥ 0. It is easy to verify that ν(φ1 φ2 ) = ν(φ1 ) + ν(φ2 ) for any non-zero
transfer functions φ1 (s) and φ2 (s). For realizability, all controller elements
IDEAL
should be proper, i.e., ν(kij ) ≥ 0. It then follows from (6.29) that

ν(Gij qrii ) ≥ ν(|G|), ∀ i ∈ m, j ∈ Ji ,

or

ν(qrii ) ≥ ν(|G|) − min ν(Gij ), ∀ i ∈ m. (6.37)

j∈Ji

Since the feedback connection of a loop transfer function does not alter its
relative degree, i.e., ν(qrii ) = ν(hri ) = 2 + νi , (6.37) is equivalent to

ν(hri ) = 2 + νi ≥ ν(|G|) − min ν(Gij ),

j∈Ji

and taking
 6.3 Achievable Performance 199

νi = max{0, ν(|G|) − min ν(Gij ) − 2} (6.38)

j∈Ji

ensures that the resultant kij are all proper and thus physically realizable.
Now all the necessary RHP zeros z, time delays Li as well as νi , required
by (6.27), have been determined. They are all feedback invariant. What re-
mains to fix are the parameters ωni and ξi in (6.27). ξi determines the reso-
nant peak magnitude Mri of the frequency response of the closed-loop transfer
function hri . The magnitude of the resonant peak gives an useful indication
of the relative stability of the system (Ogata, 1997). A large resonant peak
magnitude indicates the presence of a pair of dominant closed-loop poles with
small damping ratio, which will yield an undesirable transient response. A
small resonant peak magnitude, on the other hand, indicates the absence of
a pair of dominant closed-loop poles with small damping ratio, meaning that
the system is well damped. By ignoring the term ( 1 1s+1)νi in the operat-
Ni ωni
ing bandwidth, the magnitude of the ith closed-loop transfer function hri in
(6.27) can be calculated as
1
|hri (jω)| = q , (6.39)
ω2
(1 − 2
ωni
)2 + (2ξi ωωni )2

and for 0 ≤ ξi ≤ 0.707, its maximum value Mri is given by √1 . For

2ξi 1−ξi2
ξi > 0.707, there is no resonance, or the value of Mri is unity for ξi > 0.707.
We choose ξi to meet the specification that the closed-loop resonant peak
magnitude is less than Mri (where Mri > 1), i.e.,
1
p < Mri . (6.40)
2ξi 1 − ξi2

A ξ value of 0.707 is widely used as it results in acceptable overshoot for a

step change in the reference input and quite good damping. The overshoot
becomes excessive for values of ξi < 0.4 (Ogata, 1997).
The parameter ωni in (6.27) is directly related to ωbi , the bandwidth of
the feedback loop, which is defined as the lowest frequency such that
√
|hri (jωbi )| = |hri (0)|/ 2,

and is given by
rq
ωbi ≈ ωni (2ξi2 − 1)2 + 1 − (2ξi2 − 1)

for hri (s) in (6.27). The loop bandwidth ωbi is usually close to the 0 dB gain
cross-over frequency ωgi , at which |qrii (jωgi )| = 1. Maciejowski (1989) shows
that for typical, acceptable designs we can estimate the loop bandwidth in
terms of the cross-over frequency by
200 6. Transfer Matrix Approach

ωgi ≤ ωbi ≤ 2ωgi , (6.41)

and thus, we have
βi ωgi
ωni ≈ qp , (6.42)
(2ξi2 − 1)2 + 1 − (2ξi2 − 1)

where βi ∈ [1, 2] with a default of 1.5. Essentially, to get the benefits of

feedback control we want the loop gain, |qrii (jω)|, to be large within certain
frequency range. However, due to time delays, RHP zeros, unmodeled high
frequency dynamics and limits on the manipulated variables, the loop gain
has to drop below one at and above some frequency ωgi . Astrom (1999)
derived the following inequality for the gain cross-over frequency ωg to meet
arg Pnmp (jωg ) ≥ −π + φm − ng π/2,
where Pnmp is the non-minimum phase part of the process, φm is the required
phase margin and ng is the slope of the open loop at the cross-over frequency.
It is generally accepted that a desired loop shape for |qrii (jω)| should typi-
cally has a slope of about −1 around the desired cross-over frequency, with
preferably a steeper slope before and after the cross-over. Notice that the
Q  ni (z)
non-minimum phase part of the ith loop is given by e−Li s z∈Zi z−s z+s ,
thus we require
X  
z − jωgi
ni (z) arctan − Li ωgi ≥ −π/2 + φmi , (6.43)
z + jωgi
z∈Zi

and a straightforward method to assess the cross-over frequencies that can

be obtained for ith loop is simply to plot the left hand side of (6.43) and
determine when the inequality holds. And subsequently, ωni can be deter-
mined from (6.42). One notes that for φmi = 350 , ωgi becomes 1/L if there
is only a time delay L, or z/2 if there is a single real RHP zero at z. And for
ξi = 0.707, the parameter ωni becomes βi /Li or, βi z/2.
To consider the performance limitation imposed by input constraints, we
use a frequency-by-frequency analysis. Assume that the inputs have to meet
|ui (jω)| ≤ umax,i at each frequency and that the model has been scaled such
that the reference signal in each channel satisfies |ri (jω)| ≤ 1. For a refer-
ence signal in the ith channel, r = [0, · · · , ri , · · · , 0]T , the ideally controlled
variable y would be y = [0, · · · , hri ri , · · · , 0]T , and thus the corresponding
manipulated variable becomes
u = G−1 y = G−1 [0, · · · , hri ri , · · · , 0]T ,
i.e., elementwise as
Gij
uj = hri ri .
det(G)
 6.4 Design Procedure 201

For the worst case references (|ri (jω)| = 1), we requires

ü ü
ü Gij ü
ü ü
ü det(G) hri ü ≤ umax,j , j ∈ Ji ,

or
ü ü
ü det(G) ü
|hri | ≤ min{üü ü umax,j }.
j∈Ji Gij ü

For ω = 0, the above yields minj∈Ji {| det(G(0)) −1

Gij (0) |} > umax,j , which is necessary
for tracking a unity step reference in the ith channel with zero steady state
error. To derive an inequality on ωni imposed by input constraints, let ω = ωni
and notice from (6.39) that hri (jωni ) ≈ 2ξ1i , we require

1 |G(jωni )|
≤ min{| ij |umax,j }. (6.44)
2ξi j∈Ji G (jωni )

A straightforward method to determine the achievable ωni for ith loop under
the given input constraints is simply to plot the right hand side of (6.44).

6.4 Design Procedure

In principle, once the objective loop performance, i.e., qrii (s), is determined,
the decoupling controller K(s) can be calculated from (6.25) and (6.26). This
ideal controller decouples the process with the closed-loop transfer function
for each decoupled subsystem specified in (6.27). For a multivariable multi-
delay process in (6.6), the controller so found is usually highly complicated
and difficult to implement. We have to apply model reduction to find a much
simpler yet good approximation to the above theoretical controller. It should
be also pointed out that for a general multivariable process in (6.6), the
determination of the objective loop transfer function based on the exact |G|
and Gij (s) is actually very much involved. Therefore, the goal of this section
is, with the aid of model reduction, to ease the computation effort involved
in the determination of loop transfer function and more importantly, seek
simplest possible controllers of general rational transfer function plus delay
form, which can approximate the ideal controller such that user-specified
requirements on loop transfer function and decoupling are met.

Model Reduction. Consider a scalar system of possibly complicated

dynamics with available transfer function φ(s) or frequency response φ(jω).
The problem at hand is to find a nth-order rational function plus dead time
model:
bn sn + · · · + b1 s + 1 −Ls
φ̂(s) = φ̂0 (s)e−Ls = e , L ≥ 0, (6.45)
an sn + · · · + a1 s + a0
202 6. Transfer Matrix Approach

such that the approximation error e defined by

1 X üü ü2
N
ü
e= üW (jωk )(φ̂(jωk ) − φ(jωk ))ü , (6.46)
N
k=1

is minimized, where W (jωk ) serves as the weightings. Two very effective al-
gorithms with stability preservation have been presented in Chapter 3, one
using recursive least squares and another using the combination of the FFT
and step response. In what follows in this section, we refer “the model reduc-
tion algorithm” to one of them, or anyone else which can produce a reliable
reduced-order model in the form of (6.45).
To find the objective loop transfer function numerically, we apply the
model reduction algorithm to |G| and Gij to obtain the corresponding
reduced-order models (here, the weighting can be chosen as unity, i.e.,
W (jω) = 1). To get a model with lowest order and yet achieve a speci-
fied approximation accuracy, we start the model reduction algorithm with a
model of order 2 and then increase the order gradually to find the smallest
integer n such that the corresponding maximum relative error is less than a
user-specified  (usually 5% ∼ 15%), i.e.,
ü ü
ü φ̂(jω ) − φ(jω ) ü
ü k k ü
max ü ü ≤ . (6.47)
k=1,··· ,N ü φ(jωk ) ü

Algorithm 6.4.1 Determine the objective loop performances qrii (s) given
G(s), Mri , φmi and umax,i .
(i) Apply the model reduction algorithm to |G| and Gij , i, j ∈ m, to obtain
the corresponding reduced-order models. Determine the set Z+ |G| . Let i =
1.
(ii) Find, from the corresponding reduced-order models, τ (Gji ) and ν(Gji )
for all j ∈ m, τ (|G|) and ν(|G|) as well as η z (Gij ) for all j ∈ m and
η z (|G|) for each z ∈ Z+
|G| .
(iii) Obtain Li from (6.32), ni (z) from (6.36) for each z ∈ Z+ |G| , and νi from
(6.38). Take Ni = 10 ∼ 20.
(iv) Find the largest ωni and the smallest ξi to meet (6.40), (6.42), (6.43)
and (6.44).
(v) Form hri using (6.27) and (6.35). Calculate qrii from (6.24). Let i =
i + 1. If i ≤ m, go to step (ii); otherwise end.
After qrii , i = 1, · · · , m, are determined, the ideal controller elements
IDEAL
kji can be calculated from (6.25) and (6.26). Again, they are of compli-
cated form and it is desirable to approximate them with a simpler controller
K(s) = [kij ]. Let GK = Q = [qij ] be the actual open loop transfer function
matrix. Let the frequency range of interest be
Dωi = {s ∈ C |s = jω, − ωmi < ω < ωmi } , (6.48)
 6.4 Design Procedure 203

where ωmi > 0 is chosen to be large enough such that the magnitude of qrii
drops well below unity, or roughly, ωmi is well beyond the system bandwidth.
To ensure the actual loop to be close enough to the ideal one, we make

|qii (s) − qrii (s)|

≤ εdi < 1, s ∈ Dωi , (6.49)
|q (s)|
P rii
j6=i |qji (s)|
≤ εoi < 1, s ∈ Dωi , (6.50)
|qii (s)|

where εoi and εdi are the required performance specifications on the loop
interactions and loop performance, respectively. The values of εdi and εoi are
recommended to be in the range of 10% ∼ 30%. Smaller values of εdi and
εoi give more stringent demand on the performance and generally result in
higher order controller. One notes that although (6.50) only limits the sys-
tem couplings to a certain extent, it will still lead to almost decoupling if
the number of inputs/outputs is large. Roughly, if the number is 10, and to-
tal couplings from all other loops are limited to 30%, then each off-diagonal
element will have relative gain less than 3%, so the system is almost decou-
pled. This will be confirmed in the later simulation where almost decoupled
response is usually obtained with a εoi = 20%.
The left hand sides of (6.49) and (6.50) can be bounded respectively as
Pm m
|qii − qrii |
IDEAL
| j=1 gij (kji − kji )| X gij IDEAL
= ≤ | (kji − kji )|.
|qrii | |qrii | j=1
qrii

and
P P Pm IDEAL
j6=i |qji | j6=i | l=1 gjl (kli − kli )|
=
|qii | |qii |
Pm P IDEAL
j=1 l6=i lj (kji − kji
|g )|
≤
|qii |
ü  ü
ü
m X  ü
X ü g ü
= ü lj IDEAL ü
| | (kji − kji )ü .
ü qii 
j=1 ü l6=i ü

One notes that (6.49) implies |qii | ≥ (1 − εdi )|qrii |, and thus, (6.49) and
(6.50) are both satisfied if for each s ∈ Dωi , there hold

X m
gij (s) IDEAL
| (kji (s) − kji (s))| ≤ di ,
j=1
qrii (s)
ü  ü
m üX
X  ü
ü glj (s) ü
ü | | (kji (s) − kjiIDEAL
(s))üü ≤ (1 − di )oi ,
ü qrii (s) 
ü
j=1 l6=i ü
204 6. Transfer Matrix Approach

which can be combined into

m
X
IDEAL
|Wji (s)(kji (s) − kji (s))| ≤ 1, (6.51)
j=1

where
 
 g (s) X glj (s) 
ij
Wji (s) , max | |, | | /|qrii (s)|, (6.52)
 di (1 − di )oi 
l6=i

is a positive real valued function.

In order to satisfy (6.51), we propose to minimize the following weighted
sum of square errors
X ü ÿ
ü2
üWji (s) kji (s) − kji
IDEAL
(s) ü , (6.53)
s∈D̃ωi

where D̃ωi is a suitably chosen subset of Dωi , and the controller elements
kji are chosen to be a rational transfer function plus a time delay in the
form of (6.45). The most important frequency range for closed-loop stability
and robustness is a decade above and below ωgi , the unity gain cross-over
frequency of the ith open loop objective transfer function qrii . Therefore,
ω
D̃ωi is chosen as the interval between ω i = 10gi and ω i = 10ωgi . In general,
the resultant controller may not be in the PID form, but this is necessary
for high performance. Note also that a possible time delay is included, and
this may look unusual. The presence of dead time in the controller could
be necessary for decoupling and it may also result from model reduction,
otherwise controller order would be even higher. Though our design needs
more calculations and a controller of high-order, only simple computations
are involved and can be made automatic. The controller designed is directly
realizable.
For actual computations, kji is found again by the model reduction al-
gorithm, with the weighting taken as W (jω) = Wji (jω) as in (6.52). When
applying the model reduction algorithm, we always set a0 = 0 in (6.45) to
impose an integrator in kji (s). Let nji be the order of the controller element
kji . To find a simplest controller, the orders of the controller elements in the
ith column, nji , j = 1, · · · , m, are determined in the following way. Firstly,
an initial estimate of nji = 2, j = 1, · · · , m, are made and then the param-
eters in kji s are calculated with the model reduction algorithm. Equations
(6.49) and (6.50) are checked. If they are satisfied, kji (s), j = 1, · · · , m are
taken as the ith column controller elements; otherwise, the following indices
are evaluated
IDEAL
ρji = max |Wji (s)(kji (s) − kji (s))|. (6.54)
s∈D̃ωi
 6.4 Design Procedure 205

and let
 ü 
ü
∗
J = ∗ ü
j ∈ m üρj i = max ρji .
∗ (6.55)
j∈m

We increase nji for each j ∈ J∗ by 1 and re-calculate the corresponding kji of

increased order nji and check (6.49) and (6.50) again. This procedure repeats
until (6.49) and (6.50) are satisfied.
Algorithm 6.4.2 Find the ith column controller approximations kji (s), j =
1, · · · , m to meet (6.49) and (6.50), given G(s), qrii (s), di and oi .
ω
(i) Set the frequency range D̃ωi as [ 10gi , 10ωgi ], where ωgi is the gain cross-
IDEAL
over frequency of qrii . Evaluate kji (s), j = 1, · · · , m, from (6.25)
and (6.26).
(ii) Calculate Wji (s), j = 1, · · · , m, from (6.52). Set nji = 2, j = 1, · · · , m
and J∗ = m.
(iii) For each j ∈ J∗ , find nji -th order approximation, kji (s), to kji IDEAL
(s)
with the model reduction algorithm.
(iv) Check if (6.49) and (6.50) are satisfied. If yes, then stop; otherwise,
calculate ρji for j ∈ m from (6.54) and find J∗ from (6.55). Set nji =
nji + 1 for each j ∈ J∗ . Go to Step (iii).
Remark 6.4.1. An estimate of the time delay in kji (s) can be obtained as
follows. One notes from (6.29) that
IDEAL
τ (|G|) + τ (kji ) = τ (Gji ) + Li .
By (6.32), we have
IDEAL
τ (kji ) = τ (Gji ) − τi , (6.56)

where estimates of τ (Gji ) and τi have already been available from Algorithm
6.4.1. Simulation shows that (6.56) usually gives a very good estimate to the
delay in kji and the search on the delay can thus be avoided.
Remark 6.4.2. It is noted that model reduction on the ideal controller is
performed over D̃ωi , which is spanned over a decade below and above the
gain cross-over frequency of qrii . The magnitude and phase of the roll-off
term ( Ni1ωni s + 1)νi in (6.27) usually remains almost unchanged over that
frequency range. Hence, although it is theoretically necessary to be included
in (6.27), it has little influence in our practical design.
The entire controller design procedure is carried out by first applying
Algorithm 6.4.1 to find the objective loop transfer function and subsequently
applying Algorithm 6.4.2 to get all the controller columns. It is recaptured
as follows for ease of reference.

Overall Procedure. –Seek a single-loop controller K(s) to meet the user-

specified Mri , φmi , umax,i , di and oi , given a stable m × m process G(s).
206 6. Transfer Matrix Approach

(i) Apply the model reduction algorithm to |G| and Gij , i, j ∈ m, to obtain
the corresponding reduced-order models. Determine the set Z+ |G| . Let
i = 1.
(ii) Find, from the corresponding reduced-order models, τ (Gji ) and ν(Gji )
for all j ∈ m, τ (|G|) and ν(|G|), as well as η z (Gij ) for all j ∈ m and
η z (|G|) for each z ∈ Z+|G| . Obtain Li from (6.32), ni (z) from (6.36) for
+
each z ∈ Z|G| , and νi from (6.38). Take Ni = 10 ∼ 20.
(iii) Find the largest ωni and the smallest ξi to meet (6.40), (6.42), (6.43) and
(6.44). Form hri using (6.27) and (6.35). Calculate qrii from (6.24).
ω
(iv) Set the frequency range D̃ωi as [ 10gi , 10ωgi ], where ωgi is the gain cross-
IDEAL
over frequency of qrii . Evaluate kji (s), j = 1, · · · , m, from (6.25)
and (6.26). Calculate Wji (s), j = 1, · · · , m, from (6.52). Set nji = 2,
j = 1, · · · , m and J∗ = m.
(v) For each j ∈ J∗ , find nji -th order approximation, kji (s), to kji IDEAL
(s)
with the model reduction algorithm.
(vi) Check if (6.49) and (6.50) are satisfied. If yes, go to step (vii); other-
wise, calculate ρji for j ∈ m from (6.54) and find J∗ from (6.55). Set
nji = nji + 1 for each j ∈ J∗ . Go to Step (v).
(vii) Set i = i + 1, if i ≤ m, go to Step (ii), otherwise end the design, where
K(s) = [kij (s)] is a realizable controller designed for implementation.

6.5 Stability and Robustness Analysis

Let us consider first the nominal stability of the multivariable system in

Figure 6.1, where G(s) is not perturbed. As the process G(s) is assumed
to be stable and the controller K(s) is designed to contain no poles in the
right half of the complex plane except the origin, according to the generalized
Nyquist theorem (Chapter 3) with the Nequist Contour at the origin indented
into the right, the closed-loop system is stable if and only if the characteristic
loci of G(s)K(s) = Q(s), taken together, does not encircle the point −1.
Thus, in principle, with the generalized Nyquist theorem, the nominal
stability of the system can be determined by calculating the characteristic
loci of Q and counting their encirclements of the point −1. However, it is
not easy to calculate the characteristic loci, especially for processes of high
dimension. Notice that we are adopting a decoupling design. As a result, the
compensated open-loop transfer function Q(s) is likely to have a high degree
of column dominance, and its Gershgorin bands will be generally narrow.
By Gershgorin’s theorem (Maciejowski, 1989), we know that the union of the
Gershgorin bands “trap” the union of the characteristic loci, and we can assess
closed-loop stability by counting the encirclements of −1 by the Gershgorin
bands, if they excludes −1, since this tells us the number of encirclements
made by the characteristic loci. Thus, the closed-loop system is stable if all
the Gershgorin bands exclude and make no encirclements of the point −1.
 6.5 Stability and Robustness Analysis 207

Notice that the elements of G(s) are assumed to be strictly proper and
the elements of K(s) are designed to be proper. Thus, qji , the elements
of Q(s) = G(s)K(s) are also strictly proper, i.e., lim s→∞
Pmqji (s) = 0. Let
ωmi > 0 in (6.48) be chosen such that for all ω > ωmi , j=1 |qji (jω)| < ε
Pm
(where 0 < ε < 1), or j=1 |qji (s)| < ε for s ∈ Dωi , where Dωi denotes
Dωi ’s compliment of the Nyquist contour D, i.e., D = Dωi ∪ Dωi . We have
the following theorem.

Theorem 6.5.1. (Nominal Stability)

Pm Suppose (6.49) and (6.50) for s ∈ Dωi
and all i ∈ m, and suppose j=1 |qji (s)| < 1 and |qrii (s)| < 1 for s ∈ Dωi
and all i ∈ m,. Then the closed-loop system in Figure 6.1 is nominally stable
if
hri (s)|(di + oi + di oi ) < 1, s ∈ Dωi .

Proof. The ith Gershgorin band of Q(s), by definition, is

Gier := ∪s∈D Gier (s),

where
 ü 
 ü X 
ü
Gier (s) = z ∈ C ü |z − qii (s)| ≤ |q (s)| .
 ü ji

ü j6=i

As (6.49) and (6.50) are satisfied for s ∈ Dωi , we have for s ∈ Dωi

|qii | = |qii − qrii + qrii | ≤ |qii − qrii | + |qrii | ≤ (1 + di )|qrii |,

and
X
|qii − qrii | + |qji | ≤ di |qrii | + oi |qii | ≤ (di + oi + oi di )|qrii |.
j6=i
(6.57)

Thus for z ∈ Gier (s), where s ∈ Dωi , we have

X
|z − qrii | ≤ |qii − qrii | + |z − qii | ≤ |qii − qrii | + |qji | ≤ (di + oi + oi di )|qrii |.
j6=i

If the condition in the theorem holds, then

|1 + qrii | − |z − qrii | ≥ |1 + qrii | − (oi + di + oi di )|qrii | > 0,

which means that the distance from qrii to the point −1 is greater that from
qrii to any point in Gier so that the Nequist curve of qrii and the ith Gersh-
gorin band Gier have the same encirclememts with respect to the point −1.
208 6. Transfer Matrix Approach

As for s ∈ Dωi and i ∈ m, we have

X
|z| − |qii | ≤ |z − qii | ≤ |qji |, z ∈ Gier (s),
j6=i
Pm
and with j=1 |qji (s)| < 1, it gives
X
|z| ≤ |qii | + |qji | < 1,
j6=i

which implies that no encirclement can be made by the Gershgorin band. The
same can be said for qrii , due to the assumed |qrii (s)| < 1 for s ∈ Dωi . Thus,
the Gershgorin bands make the same number of encirclement of the point −1
as Qr = diag{qrii } does, and the system is stable as Hr = diag{hri } is so. 

In the real world where the model cannot describe the process exactly,
nominal stability is not sufficient. Robust stability of the closed-loop has to
be addressed. The multivariable control system in Figure 6.1 is referred to as
being robustly stable if the closed-loop is stable for all members of a process
family. Let the actual process be denoted as G̃(s) and the nominal process
still be denoted as G(s). The following assumptions are made.
Assumption 6.5.1 The actual process Ĝ(s) and the nominal process G(s)
do not have any unstable poles, and they are both strictly proper.
Q
Consider the family of stable processes with elementwise norm-bounded
uncertainty described by
Y  ü 
ü ĝij (jω) − gij (jω)
ü
= Ĝ = [ĝij ] ü | | = |∆ij (jω)| ≤ γij (ω) , (6.58)
gij (jω)

where γij (ω) is the bound on the multiplicative uncertainty ∆ij .

Let the perturbed open-loop transfer matrix Ĝ(s)K(s) be denoted as
Q̂(s) and its elements by q̂ji (s). As the elements of K(s) are designed to be
proper, q̂ji , are thus strictly proper, i.e., lim s→∞ q̂ji (s) =P0. Let ωmi > 0 be
m
a sufficiently large number such that for all ω > ωmi , j=1 |q̂ji (jω)| < ,
Pm
(where 0 <  < 1)) or j=1 |q̂ji (s)| <  for s ∈ Dωi . We have the following
theorem.

Theorem 6.5.2. (Robust Stability) Pm Suppose nominal stability of the closed-

loop system in Figure 6.1 and j=1 |q̂ji (s)| < 1 for s Q ∈ Dωi and all i ∈ m.
Then the system is robustly stable for all processes in if
P
l,j∈m |γlj glj kji |
|hri (s)|(di + oi + di oi ) + < 1, s ∈ Dωi .
|1 + qrii |
 6.5 Stability and Robustness Analysis 209

i
Proof. The ith Gershgorin bands Ĝer of the actual open loop transfer matrix
Q̂(s) is,
i
Ĝer = ∪s∈D Ĝier (s),
where
 ü 
 ü 
ü X
Ĝier (s) = z ∈ C ü |z − q̂ii (s)| ≤ |q̂ji (s)|,
 ü 
ü j6=i

For z ∈ Ĝier (s), where s ∈ Dωi , we have

X m
XX
|z − q̂ii | ≤ | q̂ji | = | (1 + ∆jl )gjl kli |
j6=i j6=i l=1
m
XX m
XX
≤| gjl kli | + | ∆jl gjl kli |
j6=i l=1 j6=i l=1
X m
XX
= |qji | + | ∆jl gjl kli |,
j6=i j6=i l=1

and
|z − q̂ii | ≥ |z − qrii | − |q̂ii − qrii |
Xm
= |z − qrii | − | (1 + ∆il )gil kli − qrii |
l=1
m
X
= |z − qrii | − |qii + ∆il gil kli − qrii |
l=1
m
X
≥ |z − qrii | − |qii − qrii | − | ∆il gil kli |.
l=1

Thus, we have
X X
|z − qrii | ≤ |qii − qrii | + |qji | + |∆il gil kli |.
j6=i i,j∈m

It follows from (6.57) that

X
|z − qrii | ≤ (di + oi + oi di )|qrii | + |∆il gil kli |,
i,j∈m

and
|z + 1| = |(1 + qrii ) + (z − qrii )| ≥ |1 + qrii | − |z − qrii |
 
 X 
≥ |1 + qrii | − (oi + di + oi di )|qrii | + |∆il gil kli | ,
 
i,j∈m
210 6. Transfer Matrix Approach

which will be positive, or the portion of the Gershgorin bands Gier for all
s ∈ Dωi will exclude the point −1 if
 
 X 
|1 + qrii | − (oi + di + oi di )|qrii | + |∆il gil kli | > 0,
 
i,j∈m

or
P
i,j∈m |∆il gil kli |
|hri (s)|(di + oi + di oi ) + < 1,
|1 + qrii |

and the above will be satisfied for all perturbation |∆ij (jω)| ≤ γij (ω), i, j ∈
m, if and only if
P
i,j∈m |γil (ω)gil (jω)kli (jω)|
|hri (jω)|(di + oi + di oi ) + < 1,
|1 + qrii (jω)|
for all −ω
Pmmi < ω < ωmi .
As j=1 |q̂ji (s)| < 1, for s ∈ Dωi and i ∈ m, thus for z ∈ Ĝier (s), where
s ∈ Dωi , we have
X
|z| − |q̂ii | ≤ |z − q̂ii | ≤ |q̂ji |,
j6=i

giving
X
|z| ≤ |q̂ii | + |q̂ji | < 1,
j6=i

or

|z + 1| > 0.

Therefore, the Gershgorin bands will exclude −1 for all s on the Nyquist
contour and will not change the number of encirclement of the point −1
compared with the nominal system. The system is robustly stable. 

6.6 Why PID Is Not Adequate

PID control is widely used in industry, but could be inadequate when applying
to very complex SISO processes. A frequently encountered case of PID failure
is where dynamics of individual channels are still simple but as there are
more output variables to be controlled, the interactions existing between
these variables can give rise to very complicated equivalent loops for which
PID control is difficult to handle and more importantly, the interaction for
such a process is too hard to be compensated for by PID, leading to very bad
 6.6 Why PID Is Not Adequate 211

decoupling performance and then causing poor loop performance, too, thus
making the design fail. To control such multivariable processes, PID is not
adequate any more and higher order controller with integrator is necessary.
In this section, we will demonstrate the above observations in a great detail
with a practical example.

PID Design. Among the numerous techniques currently available

for the tuning of PID controllers for a SISO process, the Astrom-Hagglund-
based method (Astrom and Hagglund, 1984) is known to work well in many
situations. The key idea of the method is to move the critical point to a
desired position on the complex plane, A1m e−j(π−φm ) , through the parameters
of the PID controller so that a combined gain and phase margin type of
specifications are satisfied. If the PID controller is described by
1
k(s) = kc (1 + + τd s),
τi s
then the parameters are given by
ku
kc = cos φm ,
Am
q
tan φm + β4 + arctan2 φm
τd = ,
2ωu
τi = βτd ,

where ku and ωu are the ultimate gain and ultimate frequency of the process,
respectively and β is usually chosen as 4. This method is here referred to as
the SV-PID tuning.
For Multivariable PID controller design, it follows from (6.22) that g̃ii
can be regarded as the equivalent process that kii faces. kii are thus designed
with respect to the equivalent processes g̃ii with some suitable PID tuning
method, say the above SV-PID tuning. The PID parameters of the controller
off-diagonal elements are determined individually by solving (6.20) at chosen
frequencies. This method is here referred to as the MV-PID tuning.
For illustration, we will study, in details, the Alatiqi process in Example
6.1.3 with G(s) given in (6.3), and show that PID, even in its multivariable
version, is usually not adequate for complex multivariable processes for good
control performance.

1 × 1 Plant. Look first at one of 1 × 1 subsystem of (6.3):

4.09e−1.3s
g11 (s) = .
(33s + 1)(8.3s + 1)
212 6. Transfer Matrix Approach

−1

−2

−3

−4

−5
−3 −2 −1 0 1 2 3 4
P ID
Fig. 6.2. Nyquist Curves (—– g11 ; - - - g11 k11 )

Its Nyquist curve is shown in Figure 6.2 in solid line. Like the other elements
of the transfer matrix, g11 (s) is basically in the form of first or second order
plus time delay and simple PID controller is adequate to control it. Using the
SV-PID tuning with Am = 2 and Φm = π/4 (the default settings throughout
this chapter unless otherwise stated), a PID controller is obtained as
P ID 0.0741
k11 (s) = 2.8512 + + 9.1421s.
s
P ID
The resultant Nyquist curve of g11 (s)k11 (s) is depicted in Figure 6.2 in
dashed line and the step response of the closed-loop system is given in Figure
6.3, showing that a PID controller is sufficient for this process and there will
not be much benefits gained by using a more complex controller.

2 × 2 Plant. We now look at one of the 2 × 2 subsystem G2×2 (s),

obtained by taking the first and second rows and columns of G(s), i.e.,
 −1.3s −0.2s

4.09e −6.36e
(33s+1)(8.3s+1) (31.6s+1)(20s+1)
G2×2 (s) =  .
−4.17e−4s 6.93e−1.01s
45s+1 44.6s+1

The equivalent processes are

g12 g21 4.09e−1.3s 3.83(44.6s + 1)e−3.19s
g̃11 = g11 − = − .
g22 (33s + 1)(8.3s + 1) (20s + 1)(31.6s + 1)(45s + 1)
 6.6 Why PID Is Not Adequate 213

1.2

0.8

0.6

0.4

0.2

−0.2
0 50 100 150

Fig. 6.3. Step Response for 1 × 1 System

and
g21 g12 6.93e−1.01s 6.84(33s + 1)(8.3s + 1)e−2.9s
g̃22 = g22 − = − .
g11 44.6s + 1 (31.6s + 1)(20s + 1)(45s + 1)
Their Nyquist curves are shown in Figures 6.4(a) and 6.4(b), respectively.
The SV-PID tuning gives the corresponding PID settings as

P ID 0.2890
k11 (s) = 3.1243 + + 6.7552s,
s
and
P ID 0.2825
k22 (s) = 1.0363 + + 0.7602s,
s
respectively. The resultant Nyquist curve of the loop transfer function g̃11 (s)
P ID P ID
k11 (s) and g̃22 (s)k22 (s) are shown in Figures 6.4(c) and 6.4(d), respec-
tively.
The MV-PID tuning yields
" #
P ID
3.1243 + 0.2890
s + 6.7552s −0.7896 + 0.4383
s + 0.5459s
K (s) = ,
1.7719 + 0.1723
s − 2.9869 1.0363 + 0.2825
s + 0.7602s

and the step response of the closed-loop system with this multivariable PID
controller is shown in Figure 6.5 with dashed lines. One notes that for this
2 by 2 process, the equivalent processes g̃11 (s) and g̃22 (s), though no longer
in simple first/second order plus delay form, is still in the capability of PID
control. However, one also notes that the decoupling from loop 1 to loop
2 is far from satisfactory, and this is because the Nyquist curve of a PID
controller is a vertical line in the complex plane and usually not sufficient to
compensate for the interactions.
214 6. Transfer Matrix Approach

0.6

0.4
0.4

0.2
0.2

0
0

−0.2
−0.2

−0.4 −0.4

−0.6 −0.6

−0.8 −0.8
−0.2 0 0.2 0.4 0.6 0.8 1 −0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

(a) Nyquist curve of g̃11 (b) Nyquist curve of g̃22

0.4 1

0.3

0.5
0.2

0.1
0
0

−0.1
−0.5

−0.2

−0.3
−1

−0.4

−0.5 −1.5
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 −1.5 −1 −0.5 0 0.5 1

P ID P ID
(c) Nyquist curve of g̃11 k11 (d) Nyquist curve of g̃22 k22

Fig. 6.4. Nyquist Curves

With the method presented in this chapter, a general controller is designed

for this 2 by 2 process as K HIGH (s) = [kij
HIGH
(s)], where

HIGH 10210s3 + 1625s2 + 93.42s + 1

k11 (s) = ,
841.2s3 + 362.7s2 + 0.4593s

HIGH −231.6s3 + 323.7s2 + 50.83s + 1

k12 (s) = ,
458.3s3 + 349.8s2 + 1.798s

HIGH 10070s3 + 1584s2 + 90.23s + 1 −2.96s

k21 (s) = e ,
1420s3 + 590.3s2 + 0.6502s

HIGH −543.7s3 + 777.7s2 + 60.68s + 1 −1.089s

k22 (s) = e ,
710.4s3 + 541.1s2 + 2.826s
and the corresponding step response is shown in Figure 6.5 with solid line.
Compared with the PID performance, the high-order controller gives much
 6.6 Why PID Is Not Adequate 215

1.5
y1
1

0.5

−0.5
0 50 100 150 200 250 300

1.5

y2 1

0.5

−0.5
0 50 100 150 200 250 300

Fig. 6.5. Step Response for 2 × 2 System (—– Proposed; - - - MV-PID tuning;)

better loop and decoupling performance. According to our experience, mul-

tivariable PID controller can yield acceptable performance for many 2 by 2
industrial process, like this example. But yet, a controller of general type, if
properly designed, can give tighter loop and decoupling performance.

3 × 3 Plant. We now proceed to a 3 × 3 subsystem obtained by deleting

the 4th row and 4th column in G(s) as
 −1.3s −0.2s −0.4s

4.09e −6.36e −0.25e
 (33s+1)(8.3s+1) (31.6s+1)(20s+1) 21s+1

 −4s
6.93e−1.01s −0.05e−5s 
G3×3 (s) =  −4.17e
45s+1 44.6s+1 34.5s+1 .
 
−1.73e−17s 5.11e−11s 4.61e−1.02s
13s+1 13.3s+1 18.5s+1

To save space, we only pick up the second loop for illustration. The equivalent
process that k22 (s) faces is now
216 6. Transfer Matrix Approach

ü ü
ü 4.09e−1.3s −6.36e−0.2s −0.25e−0.4s ü
ü (33s+1)(8.3s+1) (31.6s+1)(20s+1) 21s+1 ü
ü ü
ü −4.17e−4s 6.93e −1.01s
−0.05e −5s ü
ü 34.5s+1 üü
ü 45s+1 44.6s+1
ü −1.73e−17s −11s ü
4.61e−1.02s ü
|G| ü 5.11e
13s+1 13.3s+1 18.5s+1
g̃22 (s) = 22 = ü ü ,
|G | ü 4.09e−1.3s −0.4s ü
ü (33s+1)(8.3s+1) −0.25e21s+1 ü
ü ü
ü −1.73e−17s 4.61e−1.02s üü
ü 13s+1 18.5s+1

or,

6.93e−1.01s 5.327e−15.4s
g̃22 (s) = +
44.6s + 1 (21s + 1)(45s + 1)(13.3s + 1)

0.55e−19.88s
−
(31.6s + 1)(20s + 1)(34.5s + 1)(13s + 1)

1.045e−14.98s
+
(33s + 1)(8.3s + 1)(34.5s + 1)(13.3s + 1)

122.3e−2.9s
−
(31.6s + 1)(20s + 1)(45s + 1)(18.5s + 1)
 −1
18.85 0.432e−15.08s
· −
(33s + 1)(8.3s + 1)(18.5s + 1) (21s + 1)(13s + 1)
and its Nyquist curve is shown in Figure 6.6(a). The SV-PID tuning gives

P ID 0.4006
k22 (s) = 1.4929 + + 1.3907s.
s
P ID
The resultant loop Nyquist curves of g̃22 (s)k22 (s) is shown in Figure 6.6(b),
P ID
from which we find that this controller k22 (s) fails to stabilize the equivalent
process g̃22 (s). Notice that in this case, the Nyquist curves of the equivalent
process and loop transfer function are quite different from those of usual SISO
processes and they can be improper, which means that the magnitude of the
frequency response will increase and the Nyquist curve will not approach
to the origin as the frequency goes to infinity. In the proposed method, an
additional roll-off rate will be provided based on the process characteristic
such that the loop transfer function becomes proper and approximates an
objective transfer function. Obviously, these features cannot be obtained from
PID controller. More importantly, it is the complexity of the Nyquist curve of
the equivalent process over the medium frequency range that renders the PID
controller to be inadequate, as it is generally difficult to shape the Nyquist
curve of the equivalent process satisfactorily around this most important
frequency range using a PID controller.
Even if one could be satisfied with such a loop performance, PID would be
found to be yet inadequate to compensate for the interactions whose dynamics
 6.6 Why PID Is Not Adequate 217

0.6 2

1.5
0.4

1
0.2
0.5

0
0

−0.2 −0.5

−1
−0.4

−1.5
−0.6
−2

−0.8
−2.5

−1 −3
−1 −0.5 0 0.5 1 1.5 2 −3 −2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2

P ID
(a) Nyquist curve of g̃22 (b) Nyquist curve of g̃22 k22

25

20 −1

−2
15
−3

10 −4

−5

5
−6

−7
0
−8

−5 −9
−4 −3 −2 −1 0 1 2 3 4 5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2

IDEAL
(c) Nyquist curves (—– k32 ;-- (d) Nyquist curves (—– qr22 ; - - -
P ID HIGH
- k32 ) g̃22 k22 )

Fig. 6.6. Nyquist Curves

may also be very complicated and it would fail to achieve the decoupling. For
an illustration, according to (6.20), the controller off-diagonal element, say
k32 (s), should be
ü ü
ü 4.09e−1.3s −6.36e−0.2s ü
ü (33s+1)(8.3s+1) (31.6s+1)(20s+1) ü
ü ü
ü −1.73e−17s 5.11e −11s ü
ü 13s+1 13.3s+1
ü
IDEAL
k32 (s) = − üü −1.3s
ü k22 (s),
−0.4s ü
(6.59)
4.09e −0.25e
ü (33s+1)(8.3s+1) 21s+1 ü
ü ü
ü −1.73e−17s 4.61e−1.02s üü
ü 13s+1 18.5s+1

P ID
with k22 (s) = k22 (s). The MV-PID design method gives the PID form of
P ID
the off-diagonal controller elements k32 (s) as
218 6. Transfer Matrix Approach

P ID 0.2152
k32 (s) = −1.6825 − − 2.381s.
s
IDEAL
The Nyquist curve of the ideal decoupler k32 (s) and its PID approxima-
P ID
tion k32 are shown in Figure 6.6(c) in solid line and dash line, respectively,
from which one sees that the Nyquist curve of the ideal decoupler k32 (s)
P ID
is quite complicated and thus its approximation by k32 (s) of PID form is
inadequate. Actually, with the MV-PID tuned controller:
 
5.976 + 0.76 0.61
s + 11.7s 2.16 + s + 2.72s −0.825 +
0.218
s + 0.994s
 
K P ID (s) =  0.456 0.41 0.133
 2.59 + s − 5.87s 1.49 + s + 1.39s 0.56 + s + 0.328s  ,


−9.25 − 0.22 0.22

s + 5.8s −1.68 − s − 2.4s 1.169 +
0.194
s + 1.76s
the step response for the closed-loop system shown in Figure 6.7 with dash
lines has gone unstable.

1.5
y1
1

0.5

0
0 100 200 300 400 500 600

1.5
y2
1
0.5
0
−0.5
0 100 200 300 400 500 600
2

y3
1

−1
0 100 200 300 400 500 600

Fig. 6.7. Step Response for 3 × 3 System (—– Proposed; - - - MV-PID)

Generally, an equivalent process in a multivariable process can become

highly complicated as the number of inputs and outputs increases. The
Nyquist curve can exhibit strange shape especially around the cross-over
frequency and thus PID controllers, which have only three adjustable pa-
rameters, are difficult to shape it satisfactorily or even fail to stabilize the
 6.6 Why PID Is Not Adequate 219

equivalent processes. Of course, One can always de-tune the PID controllers,
or reduce the gains to a sufficient degree to generate a stable closed-loop if
the original process is stable. However, we have to have the sluggish response
and poor performance, which might not be desirable or acceptable.
With our method, an objective loop transfer function for the second loop
for this example is specified as

0.17902 e−6.62s
qr22 (s) = ,
s2 + 2 × 0.1790 × 0.707s + 0.17902 (1 − e−6.62s )
HIGH
and a 6th-order controller, k22 (s), is designed as

HIGH 7266s6 − 3613s5 + 1754s4 − 1061s3 + 626.7s2 + 60.79s + 1

k22 (s) = ,
59390s6 + 35180s5 + 22460s4 + 5731s3 + 1458s2 + 8.781s

HIGH
such that g̃22 (s)k22 (s) approximates qr22 (s). The Nyquist curve of the
IDEAL
actual loop transfer function g̃22 (s)k22 (s) and the ideal one qr22 (s) are
shown in Figure 6.6(d) with dashed line and solid line, respectively. We can
HIGH HIGH
see that g̃22 (s)k22 (s) with the 6th-order k22 (s) yields very close fitting
to the ideal loop transfer function in frequency domain while generally PID
controller is not capable of that.
To achieve decoupling, according to (6.20), the controller off-diagonal el-
HIGH
ement, say k32 (s), should be given by (6.59) with k22 (s) = k22 (s). The
IDEAL
Nyquist curve of the ideal decoupler k32 (s) is shown in Figure 6.8(a)
with solid line, which is again a function that PID is difficult to approximate.
HIGH
A 6th-order approximation k32 (s) is obtained as

HIGH 2110s5 − 3671s4 + 631s3 + 1756s2 + 68.43s + 1 −12.35s

k32 (s) = e ,
−39970s5 − 19950s4 − 8978s3 − 2266s2 − 17.41s

and its Nyquist curve is shown in Figure 6.8(a) with dashed line, from which
we see that it approximates the ideal one very well over the interested fre-
quency range. The other controller elements are obtained as
HIGH −4433s4 − 1021s3 + 679.6s2 + 50.6s + 1
k11 (s) = ,
10580s4 + 2104s3 + 1006s2 + 6.468s

HIGH 373s6 − 262s5 + 153s4 + 875s3 + 260s2 + 50.8s + 1 −2.03s

k12 (s) = e ,
55890s6 + 19200s5 + 16960s4 + 2820s3 + 943s2 + 5.8s

HIGH −5094s4 − 351.7s3 + 501.5s2 + 45.67s + 1

k13 (s) = ,
231600s4 + 47270s3 + 17520s2 + 110.4s

HIGH −4062s4 − 1142s3 + 666.5s2 + 49.89s + 1 −2.86s

k21 (s) = e ,
17660s4 + 3408s3 + 1675s2 + 10.8s
220 6. Transfer Matrix Approach

HIGH −1820s3 + 752.3s2 + 43.82s + 1 −3.492s

k23 (s) = e ,
118900s3 + 27330s2 + 184.3s

HIGH −19920s4 + 3943s3 + 1099s2 + 60.08s + 1 −13.94s

k31 (s) = e ,
−31440s4 − 6909s3 − 3508s2 − 22.31s

HIGH −24880s3 + 7571s2 + 434.8s + 1 −5.085s

k33 (s) = e .
82100s3 + 43020s2 + 137.4s

The time domain step response of the overall multivariable feedback system
is shown in Figure 6.7 with solid lines, from which we see that the general
controller gives very good loop and decoupling performance.

0.1
5

0
4
−0.1
3
−0.2

2
−0.3

1
−0.4

0
−0.5

−1 −0.6

−2
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 0 0.5 1 1.5

IDEAL
(a) Nyquist curves (—– k32 ;- (b) Nyquist curve of g̃22
HIGH
- - k32 )

2 20

1.5
0
1

0.5 −20

0
−40
−0.5

−60
−1

−1.5
−80

−2
−100
−2.5

−3 −120
−4 −3 −2 −1 0 1 2 3 4 −70 −60 −50 −40 −30 −20 −10 0 10 20

P ID IDEAL
(c) Nyquist curve of g̃22 k22 (d) Nyquist curves (—– k32 ; -
P ID
- - k32 )

Fig. 6.8. Nyquist Curves

 6.6 Why PID Is Not Adequate 221

4 × 4 Plant. Now, we look at the original 4 by 4 process G(s). Again,

we pick up the second loop for illustration. Under decoupling, the equivalent
process that k22 (s) faces is
ü ü
ü 4.09e−1.3s −6.36e−0.2s −0.25e−0.4s −0.49e−5s ü
ü (33s+1)(8.3s+1) (31.6s+1)(20s+1) 21s+1 22s+1 ü
ü ü
ü −4.17e−4s 6.93e−1.01s −0.05e−5s 1.53e−2.8s ü
ü 45s+1 44.6s+1 34.5s+1 48s+1 ü
ü ü
ü −1.73e−17s 5.11e −11s
4.61e −1.02s
−5.48e −0.5s ü
ü ü
ü 13s+1 13.3s+1 18.5s+1 15s+1 ü
ü −11.18e−2.6s −0.02s −0.05s −0.6s ü
ü 14.04e −0.1e 4.49e ü
(43s+1)(6.5s+1) (45s+1)(10s+1) (31.6s+1)(5s+1) (48s+1)(6.3s+1)
g̃22 (s) = ü ü .
ü 4.09e−1.3s −0.25e−0.4s −0.49e−5s ü
ü (33s+1)(8.3s+1) 21s+1 22s+1 ü
ü ü
ü −1.73e−17s 4.61e−1.02s −5.48e−0.5s ü
ü 13s+1 18.5s+1 15s+1 ü
ü ü
ü −11.18e−2.6s −0.05s −0.6s ü
ü (43s+1)(6.5s+1) (31.6s+1)(5s+1)
−0.1e 4.49e
(48s+1)(6.3s+1)
ü

Its Nyquist curve is shown in Figure 6.8(b), and the complexity of the equiv-
alent process is obvious. The SV-PID tuning gives the PID controller as
P ID 0.7503
k22 (s) = 3.5264 + + 4.1438s.
s
P ID1
The resultant loop Nyquist curves of g̃22 (s)k22 (s) is shown in Figure 6.8(c),
from which we can expect very poor loop performance or instability.
Apart from the poor loop performance, PID is also found to be inadequate
to achieve decoupling for this process. To achieve decoupling, (6.20) applies,
say for k32 (s), to get
ü ü
ü 4.09e−1.3s −6.36e−0.2s −0.49e−5s ü
ü (33s+1)(8.3s+1) (31.6s+1)(20s+1) 22s+1 ü
ü ü
ü −1.73e−17s 5.11e −11s
−5.48e −0.5s ü
ü 13s+1 13.3s+1 15s+1 ü
ü ü
ü −11.18e−2.6s −0.02s −0.6s ü
ü (43s+1)(6.5s+1) (45s+1)(10s+1) (48s+1)(6.3s+1) ü
14.04e 4.49e
IDEAL
k32 (s) = − ü ü k22 (s),
ü 4.09e−1.3s −0.25e−0.4s −0.49e−5s ü
ü (33s+1)(8.3s+1) 21s+1 22s+1 ü
ü ü
ü −1.73e−17s 4.61e−1.02s −5.48e−0.5s ü
ü 13s+1 18.5s+1 15s+1 ü
ü ü
ü −11.18e−2.6s −0.05s −0.6s ü
ü (43s+1)(6.5s+1) (31.6s+1)(5s+1)
−0.1e 4.49e
(48s+1)(6.3s+1)
ü
(6.60)
P ID
with k22 (s) = k22 (s). The MV-PID design method gives the PID form of
P ID
the off-diagonal controller elements k32 (s) as

P ID 1.9152
k32 (s) = −3.7326 + − 3.5149s.
s
IDEAL
The Nyquist curve of the ideal decoupler k32 (s) and its PID approxima-
P ID
tion k32 are shown in Figure 6.8(d) in solid line and dash line, respectively,
222 6. Transfer Matrix Approach

indicating invalidation of the approximation. Actually, the closed-loop sys-

tem with K(s) being the multivariable PID controller designed with MV-PID
is unstable.
With our design procedure, an objective loop transfer function for the
second loop for this plant is specified as
0.032422 e−36.54s
qr22 (s) = .
s2 + 2 × 0.03242 × 0.707s + 0.032422 (1 − e−36.54s )
HIGH
and a 4th-order k22 (s):

HIGH −146.2s4 − 121.3s3 + 16.19s2 + s + 0.009641 −25.35s

k22 (s) = e ,
15340s4 + 1519s3 + 114.8s2 + 1.144s

HIGH
is designed such that g̃22 (s)k22 (s) approximates qr22 (s). The Nyquist
HIGH
curve of the actual loop transfer function g̃22 (s)k22 (s) and the ideal loop
one qr22 (s) are shown in Figure 6.9(a) with dash line and solid line, respec-
HIGH
tively, and they are different. Note that k22 (s) contains a time delay term.
This can happen and is necessary in the multivariable case for decoupling,
as shown before. If this delay is removed or reduced, or in other word, the
time delay in the second decoupled loop transfer function qr22 is reduced,
then one or some controller off-diagonal element will have time predictions
(eLs , L > 0), which is not realizable.

1 1

0
0

−1
−1
−2

−2
−3

−3 −4

−4 −5

−6
−5

−7
−6
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4

IDEAL
(a) Nyquist curves (—– qr22 ; - - - (b) Nyquist curves (—– k32 ;-
HIGH HIGH
g̃22 k22 ) - - k32 )

Fig. 6.9. Nyquist Curves

The decoupling condition (6.20) determines the controller off-diagonal ele-

HIGH
ments. For instance, k32 (s) should be given by (6.60) with k22 (s) = k22 (s).
IDEAL
The Nyquist curve of the ideal decoupler k32 (s) is shown in Figure 6.9(b)
 6.6 Why PID Is Not Adequate 223

with solid line, which is again a function that PID is difficult to approximate.
HIGH
A 5th-order approximation k32 (s) is obtained with the proposed method
as
66110s5 − 841.8s4 − 107.1s3 + 216.1s2 + s − 0.1403 −1.683s
k32 (s) = e ,
−619800s5 − 102700s4 − 10680s3 − 545.2s2 − 14.23s

and its Nyquist curve is shown in Figure 6.9(b) with dash line, from which we
see that it approximates the ideal one very well over the interested frequency
range. The remaining controller elements generated by the proposed method
are listed in (6.61) to (6.74). The resultant time domain step response is
shown in Figure 6.10, from which we see that the general controller gives
excellent loop and decoupling performance.

y1
1
0.5
0
−0.5
0 200 400 600 800 1000 1200 1400 1600 1800 2000

y2 1
0.5
0
−0.5
0 200 400 600 800 1000 1200 1400 1600 1800 2000

y3 1
0.5
0
−0.5
0 200 400 600 800 1000 1200 1400 1600 1800 2000

y4 1
0.5
0
−0.5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Fig. 6.10. Step response for Example 6.1.3

In view of the above case study, we observe that

(i) There are many complex multivariable processes. The complexity of the
processes generally does not come from the complexity of the individual
channels, but from the multivariable interactions, which may become
224 6. Transfer Matrix Approach

very difficult to compensate for when the I/O number of the processes
increases. The resultant equivalent processes can be extremely compli-
cated even when all the individual channel dynamics are of simple first
or second order plus time delay form.
(ii) For such complex multivariable processes, PID controller is usually in-
adequate, not only because the equivalent processes is difficult to be
controlled by PID, but also the interactions are too difficult to be com-
pensated for by PID.
(iii) Existing non-PID control are usually state-space based. Their specifi-
cations are generally quite different from those used in process control.
Time delay is frequently encountered in process control while state-space
based approach usually cannot handle the time delay easily. The decou-
pling requirement which is very important for process control is usually
not considered explicitly in state-space based approaches. These explain
for discrepancy between the many well developed state-space based mul-
tivariable designs (H∞ , LQG, ...) and the rare use of them in process
control.
(iv) A general, effective and efficient design of non-PID control for com-
plex multivariable process capable of achieving high performance, is thus
needed and should have great potential for industrial applications. This
chapter has presented such a method.

6.7 More Simulation

In this section, the design method presented in Sections 2-4 is applied to the
five industrial processes described in Section 1 to show its effectiveness.
Example 6.1.1 (continued). Consider the Tyreus distillation column
in (6.1). Let ξi = 0.707, i.e., Mri = 1, and φmi = π/4, di = 0.2, oi = 0.2,
i = 1, 2, 3, the design procedure is activated.
(i) The application of the model reduction algorithm to |G| and Gij yields
the following reduced-order models:
−3724s2 + 11260s + 54.77
|G(s)| = e−15.3s ,
462800s3 + 21080s2 + 272.8s + 1
−61.22s + 30.13
G11 (s) = e−0.946s ,
490.4s2 + 44.12s + 1
−8.417s + 1.09
G12 (s) = e−11.47s ,
−169.7s2 + 13.06s + 1
3.521s − 0.1071 −14.77s
G13 (s) = e ,
532.8s2 + 45.3s + 1
−2024s2 − 27510s − 16.21
G21 (s) = e−14.22s ,
274000s3 + 20140s2 + 449.3s + 1
 6.7 More Simulation 225

230.5s + 17.25
G22 (s) = e−5.847s ,
1718s2 + 92.27s + 1
817.3s2 − 4474s − 20.48
G23 (s) = e−14.81s ,
504200s3 + 22520s2 + 288.9s + 1
12630s2 + 376.3s + 14.45
G31 (s) = e−8.905s ,
68680s3 + 8196s2 + 420.1s + 1
157.6s + 4.849
G32 (s) = e−4.334s ,
2069s2 + 97.88s + 1
−348s2 + 266.3s + 0.5485
G33 (s) = e−0.8209s .
121800s3 + 28450s2 + 478.9s + 1
We got Z+ |G| = {3.027} from |G(s)|. Set i = 1.
(ii) From the above reduced order models, we obtain τ (|G|) = 15.3,
τ (G11 ) = 0.946, τ (G12 ) = 11.47, τ (G13 ) = 14.77; for z = 3.027 in Z+
|G| ,
ηz (|G|) = 1 and ηz (G1j ) = 0, j = 1, 2, 3; τ1 = 0.9463 and L1 = 14.35.
And n1 (z) = 1.
(iii) Take ξ1 = 0.707. With φm1 = π/4, ωg1 is obtained from (6.43) as
ωg1 = 0.0529, and ωn1 is determined from (6.42) as ωn1 = 0.0794. hr1
is then formed from (6.27) as

0.07942 (3.027 − s)e−14.35s

hr1 (s) = ,
(s2 + 2 × 0.0794 × 0.707s + 0.07942 )(3.027 + s)

and qr11 is calculated from (6.24).

ωg1 IDEAL
(iv) Specify D̃ω1 with ω 1 = 10 and ω 1 = 10ωg1 . kj1 (s), j = 1, 2, 3, are
evaluated from (6.25) and (6.26). Wj1 (s), j = 1, · · · , m, are found from
(6.52). Set nj1 = 2, j = 1, 2, 3 and J∗ = {1, 2, 3}.
(v) For each j ∈ J∗ , the nj1 th-order approximation to kj1 (s) is found with
the model reduction algorithm as

−689.7s2 + 57.31s + 1
k11 (s) = ,
710.8s2 + 64.14s
20330s2 + 235.2s + 1 −10.52s
k21 (s) = e ,
319900s2 + 1400s
−50380s2 + 462.1s + 1 −13.82s
k31 (s) = e .
−4709000s2 − 74420s
(vi) It is found that (6.49) and (6.50) are not satisfied. ρ11 , ρ21 and ρ31 are
then calculated as 6.965, 6.858 and 4.400 from (6.54) and J∗ is thus
found as {1} from (6.55). Set n11 = 3, i.e., the order of k̂11 (s) increases
by 1. Go back to step (v). This procedure repeats until (6.49) and (6.50)
are both satisfied and the following k11 , k21 and k31 are obtained:
226 6. Transfer Matrix Approach

24.09s3 − 8.787s2 + s + 0.01918

k11 (s) = ,
49.83s3 + 6.991s2 + 1.214s
−69.82s3 + 10.68s2 + s + 0.01277 −10.52s
k21 (s) = e ,
1041s3 + 126.5s2 + 22.49s
537.2s3 − 120.6s2 + s + 0.07195 −13.82s
k31 (s) = e .
−55630s3 − 7997s2 − 1186s
(vii) We repeat the above process for i = 2 and find the 2nd-column of the
controller as
24260s6 − 10870s5 + 2651s4 − 110s3 + 55.22s2 + s + 0.0006456 −8.376s
k12 (s) = e ,
−57140s6 − 17230s5 − 3013s4 − 376.2s3 − 23.97s2 − 0.05722s

−1386s6 − 1555s5 + 190.9s4 + 105.1s3 + 21.64s2 + s + 0.0024

k22 (s) = ,
120100s6 + 20740s5 + 8681s4 + 717.5s3 + 94.53s2 + 0.2149s

37.42s4 − 20.83s3 + 6.675s2 + s + 0.05205 −8.962s

k32 (s) = e .
−2512s4 − 878.2s3 − 97.77s2 − 3.945s

For i = 3, we have
−1430s4 + 938.5s3 + 29.26s2 + s + 0.008427 −8.084s
k13 (s) = e ,
34560s4 + 2527s3 + 485.3s2 + 1.039s
−49.8s3 + 6.178s2 + s + 0.02832 −3.513s
k23 (s) = e ,
554.6s3 + 62.99s2 + 11.34s
−32.71s3 − 3.641s2 + s + 0.03188
k33 (s) = .
4964s3 + 411s2 + 102.3s
If we restrict the controller to the PID type to approximate the ideal
controller, the design procedure produces
P ID
K =
 0.0152 0.0475 0.0022 
−0.0442 + s + 1.2413s −0.0095 − s − 0.7917s −0.0445 + s + 0.2294s
 
 0.0062 + 0.0006
+ 0.0715s 0.0998 + 0.0107
− 0.4601s 0.0198 + 0.0021
− 0.2532s 
 s s s .
0.0356 0.0113 0.0005
−0.0000538 + 1000s + 0.0188s 0.0342 − s − 0.2112s −0.0073 + s + 0.0298s.

The step response of the closed-loop system is shown in Figure 6.11, where
the solid line is the response with K(s) and the dash line with K P ID (s), from
which we can see that the performance of the general controller is very satis-
factory while the PID performance is not. The response for the manipulated
variables are shown in Figure 6.12, from which we see that the manipulated
variables are quite small and smooth for the high-order controller while PID
control generates big peaks from the derivative terms.
In order to see the robustness of the proposed method, we increase the
static gains in all the elements in the process transfer matrix by 40%, i.e.,
each gij (s) is perturbed to 1.4gij (s). The step response under such gain per-
turbations is shown in Figure 6.13 with dashdot lines, while the nominal
 6.7 More Simulation 227

1.5
y1
1

0.5

0
0 100 200 300 400 500 600 700 800 900 1000

y2 1

0.5

0
0 100 200 300 400 500 600 700 800 900 1000

y3
1

0.5

−0.5
0 100 200 300 400 500 600 700 800 900 1000

Fig. 6.11. Output Step Responses for Example 6.1.1 (—– High order; - - - PID)

u1 1
0
−1
−2

0 100 200 300 400 500 600 700 800 900 1000

0.6
u2
0.4
0.2
0
−0.2
0 100 200 300 400 500 600 700 800 900 1000

u3 0

−0.2

−0.4

0 100 200 300 400 500 600 700 800 900 1000

Fig. 6.12. Input Step Responses for Example 6.1.1C (—– High order; - - - PID)
228 6. Transfer Matrix Approach

y1
1

0.5

0
0 100 200 300 400 500 600 700 800 900 1000

y2
1

0.5

0
0 100 200 300 400 500 600 700 800 900 1000

y3
1

0.5

−0.5
0 100 200 300 400 500 600 700 800 900 1000

Fig. 6.13. Robustness for Example 6.1.1C (—– nominal performance; − · − · −·

under gains variation; - - - under time constants variation;)

performance is shown in solid lines. One notes that this gain variation does
not affect the decoupling, which is expected. To introduce variation in the
system dynamics, we perturb the nominal process by increasing the time
constants in all gij (s) by 40%. The corresponding step response is shown in
Figure 6.13 with dashed lines, exhibiting the deterioration of both the loop
and decoupling performance. However, such deterioration is reasonable as
compared with the “size” of the perturbation.
Example 6.1.2 (continued). Consider the 4 by 4 Doukas process in
(6.2). Following our method, the objective closed-loop transfer functions are
determined as
0.044422 e−17.78s
hr1 (s) = ,
s2 + 2 × 0.04442 × 0.707s + 0.044422
0.045272 e−17.45s
hr2 (s) = ,
s2 + 2 × 0.04527 × 0.707s + 0.045272
0.045192 e−17.48s
hr3 (s) = ,
s2 + 2 × 0.04519 × 0.707s + 0.045192
0.096682 e−8.17s
hr4 (s) = .
s2 + 2 × 0.09668 × 0.707s + 0.096682
 6.7 More Simulation 229

The resulting controller is given by

−41.02s3 + 5.685s2 + 1s + 0.02819 −4.202s
k11 (s) = e ,
−565.3s3 − 63.11s2 − 11.35s
−77.03s3 + 7.328s2 + 1s + 0.01236 −9.929s
k12 (s) = e ,
−1503s3 − 147.2s2 − 25.3s
−658s6 − 2209s5 − 70.18s4 + 173.4s3 + 17.88s2 + s + 0.005104 −2.959s
k13 (s) = e ,
−138600s6 − 26980s5 − 9990s4 − 906.6s3 − 97.3s2 − 0.4822s
1.299s2 + s − 0.007045
k14 (s) = ,
103.3s2 + 18.46s
11620s3 − 1709s2 + s − 1.847
k21 (s) = ,
89460s3 + 5427s2 + 597.3s
34.76s3 − 9.855s2 + s + 0.01976 −0.3349s
k22 (s) = e ,
−63.31s3 − 8.238s2 − 1.377s
12730s6 − 3075s5 + 1485s4 − 361.3s3 + 42.57s2 + s + 0.0008608
k23 (s) = ,
84620s6 + 15440s5 + 4828s4 + 484.7s3 + 29.44s2 + 0.09661s
4.94s2 + 1s + 0.01047 −11.05s
k24 (s) = e ,
−66.5s2 − 3.176s

−20.65s3 + 4.144s2 + s + 0.028 −6.645s

k31 (s) = e ,
−3131s3 − 305.5s2 − 68.91s
1597s3 − 344s2 + 1s + 0.1779 −15.17s
k32 (s) = e ,
191800s3 + 24400s2 + 3503s
−247.2s5 + 49.82s4 + 2.14s3 + 10.2s2 + s + 0.1042 −12.71s
k33 (s) = e ,
20850s5 + 4104s4 + 1361s3 + 109.6s2 + 10.4s
−0.5092s2 + s + 0.02297 −4.705s
k34 (s) = e ,
−772.8s2 − 16.49s
−14.91s3 − 2.164s2 + s + 0.03425 −3.75s
k41 (s) = e ,
4279s3 + 453.7s2 + 90.39s
12.65s3 − 11.83s2 + s + 0.001142
k42 (s) = ,
24270s3 + 1367s2 + 388.5s
21.11s4 + 17.06s3 + 5.502s2 + s + 0.1161 −14.14s
k43 (s) = e ,
2753s4 + 446.6s3 + 150s2 + 10.8s
−1.499s2 + s + 0.08238 −2.63s
k44 (s) = e .
165.2s2 + 13.17s
230 6. Transfer Matrix Approach

The step response is shown in Figure 6.14 with superior performance.

y1 1
0.5
0
−0.5
0 200 400 600 800 1000 1200

y2 1
0.5
0
−0.5
0 200 400 600 800 1000 1200

y3 1
0.5
0
−0.5
0 200 400 600 800 1000 1200

y4 1
0.5
0
−0.5
0 200 400 600 800 1000 1200
Fig. 6.14. Step Responses for Example 6.1.2C

Example 6.1.3C (continued). Consider the 4 by 4 Alatiqi process in

(6.3). By the design procedure, the objective closed-loop transfer functions
are
0.027082 e−43.75s
hr1 (s) = ,
s2 + 2 × 0.02708 × 0.707s + 0.027082
0.032422 e−36.54s
hr2 (s) = ,
s2 + 2 × 0.03242 × 0.707s + 0.032422
0.026672 e−44.42s
hr3 (s) = ,
s2 + 2 × 0.02667 × 0.707s + 0.026672
0.025192 e−47.04s
hr4 (s) = .
s2 + 2 × 0.02519 × 0.707s + 0.025192
 6.7 More Simulation 231

And the controller is

1526s4 − 171s3 + 14.62s2 + s + 0.01487
k11 (s) = , (6.61)
19500s4 + 1616s3 + 122.2s2 + 1.879s
−18920s6 − 17290s5 + 3341s4 − 453.8s3 − 1.136s2 + s + 0.01
k12 (s) = , (6.62)
791900s6 + 247400s5 + 33380s4 + 2413s3 + 91.32s2 + 0.82s

−12.21s2 + s + 0.02426
k13 (s) = , (6.63)
1982s2 + 50.13s

−123.4s3 + 3.941s2 + s − 0.01903 −1.958s

k14 (s) = e , (6.64)
2471s3 + 114.7s2 + 7.317s

1104s4 − 107.1s3 + 10.31s2 + s + 0.02 −0.8466s

k21 (s) = e , (6.65)
145000s4 + 11950s3 + 958.6s2 + 13.55s

−13.97s2 + s + 0.04257 −3.335s

k23 (s) = e , (6.66)
20520s2 + 392s

−553.9s3 − 31.68s2 + s + 0.1363 −24.2s

k24 (s) = e , (6.67)
−24360s3 − 1085s2 − 70.19s

21.66s3 − 7.492s2 + s + 0.03177 −1.013s

k31 (s) = e , (6.68)
395.4s3 + 20.1s2 + 1.61s

296.4s3 − 23.09s2 + s + 0.1881 −3.887s

k33 (s) = e , (6.69)
13050s3 + 864.5s2 + 54.53s

−122.6s3 + 0.9794s2 + s + 0.008617 −3.358s

k34 (s) = e , (6.70)
962.4s3 + 40.13s2 + 2.606s

19.56s3 − 6.984s2 + s + 0.03433 −2.077s

k41 (s) = e , (6.71)
542s3 + 28.34s2 + 2.217s
19180s5 + 12170s4 − 2071s3 + 308.3s2 + s − 0.2772 −1.541s
k42 (s) = e , (6.72)
−478900s5 − 115500s4 − 13660s3 − 835.9s2 − 24.39s

−20.23s2 + s + 0.02799 −2.502s

k43 (s) = e , (6.73)
643.4s2 + 29.45s

−76.07s3 − 4.884s2 + s + 0.00616

k44 (s) = . (6.74)
1235s3 + 51.16s2 + 3.434s
The step response is shown in Figure 6.10.
Example 6.1.4 (continued). Consider the Ammonia process model
(Yokogawa Ltd.) in (6.4). Following the procedure, we obtain

0.20572 e−5.760s
hr1 (s) = ,
s2 + 2 × 0.2057 × 0.707s + 0.20572
0.057162 e−20.73s
hr2 (s) = ,
s2 + 2 × 0.05716 × 0.707s + 0.057162
232 6. Transfer Matrix Approach

0.13532 e−8.760s
hr3 (s) = ;
s2 + 2 × 0.1353 × 0.707s + 0.13532
and
−0.6015s2 + s + 0.09952 −2s
k11 (s) = e ,
−2.783s2 − 1.409s
3.685s3 + 4.08s2 + s + 0.08144 −1.861s
k12 (s) = e ,
5.537s3 + 0.793s2 + 0.1084s
−0.1618s3 + 1.721s2 + s + 0.1044 −2s
k13 (s) = e ,
2.406s3 + 0.8755s2 + 0.2685s
−0.0771s3 + 0.5865s2 + s + 0.1633 −3.513s
k21 (s) = e ,
−0.09085s3 − 0.06274s2 − 0.02637s
−0.5531s3 + 2.607s2 + s + 0.1366 −3.494s
k22 (s) = e ,
−0.09555s3 − 0.01373s2 − 0.00194s
4.332s3 + 2.386s2 + s − 0.1834 −3.936s
k23 (s) = e ,
0.1233s3 + 0.04899s2 + 0.01975s
−0.5639s2 + s + 0.05756
k31 (s) = ,
−1.875s2 − 1.051s
−4.014s4 + 7.842s3 + 4.624s2 + s + 0.0617
k32 (s) = ,
−7.849s4 − 5.138s3 − 0.7457s2 − 0.08274s
−0.306s3 + 2.29s2 + s + 0.07093
k33 (s) = .
2.016s3 + 0.6972s2 + 0.2153s
The step response is shown in Figure 6.15.
Example 6.1.5 (continued). Consider the depropanizer process in
(6.5). It follows from the design procedure that

0.0069982 (0.01495 − s)e−40.72s

hr1 (s) = ,
(s2 + 2 × 0.006998 × 0.707s + 0.0069982 )(0.01495 + s)

0.0067592 (0.01495 − s)e−44.38s

hr2 (s) = ,
(s2 + 2 × 0.006759 × 0.707s + 0.0067592 )(0.01495 + s)

0.0073572 (0.01495 − s)e−38.73s

hr3 (s) = ;
(s2 + 2 × 0.007357 × 0.707s + 0.0073572 )(0.01495 + s)

and
−53.59s3 + 24.86s2 + s + 0.005836
k11 (s) = ,
4217s3 + 85.86s2 + 1.091s
 6.7 More Simulation 233

y1
1

0.5

−0.5
0 100 200 300 400 500 600

y2
1

0.5

−0.5
0 100 200 300 400 500 600

y3
1

0.5

−0.5
0 100 200 300 400 500 600

Fig. 6.15. Step Responses for Example 6.1.4C

−63.61s3 + 8.084s2 + s + 0.008384 −27.47s

k12 (s) = e ,
7184s3 + 151.5s2 + 1.743s
122.6s3 − 6.025s2 + s + 0.01103
k13 (s) = e−33.82s ,
1.689 × 10+004 s3 + 333.5s2 + 3.802s
−18.72s3 + 6.678s2 + s + 0.00809 −17.67s
k21 (s) = e ,
2.37 × 10+004 s3 + 514.8s2 + 6.357s
26.31s3 + 28.53s2 + s + 0.006068 −11.81s
k22 (s) = e ,
7.887 × 10+004 s3 + 1640s2 + 20.17s
227.8s3 + 4.625s2 + s + 0.01424 −26.42s
k23 (s) = e ,
1.849 × 10+005 s3 + 3658s2 + 42.94s
27.48s3 − 2.061s2 + s + 0.008284 −0.6552s
k31 (s) = e ,
1532s3 + 32.7s2 + 0.4058s
59.87s3 − 1.332s2 + s + 0.009638 −0s
k32 (s) = e ,
4699s3 + 94.31s2 + 1.06s
101.5s3 − 9.961s2 + s + 0.002642
k33 (s) = ,
7283s3 + 154.2s2 + 1.597s
The step response is shown in Figure 6.16.
234 6. Transfer Matrix Approach

y1
1

0.5

−0.5
0 1000 2000 3000 4000 5000 6000

y2
1

0.5

−0.5
0 1000 2000 3000 4000 5000 6000

y3
1

0.5

−0.5
0 1000 2000 3000 4000 5000 6000

Fig. 6.16. Step Responses for Example 6.1.5C

6.8 Conclusions

In this chapter, a systematic design method is developed to achieve fastest

loop speed with acceptable overshoot and minimum loop interactions with a
controller of least complexity for complex multivariable processes. The fun-
damental relations for decoupling a multivariable process with time delay has
been established. All the unavoidable time delays and non-minimum phase
zeros that are inherent in a decoupled feedback loop are characterized. The
objective loop transfer functions are then suitably specified to achieve fastest
possible response taking into account the performance limitation imposed by
the unavoidable non-minimum phase zeros and time delays. The ideal con-
troller is obtained and it is generally a complicated irrational transfer matrix,
for which model reduction is applied to fit it as a much simpler transfer ma-
trix with its elements in rational transfer function plus delay form. To obtain
a controller of least complexity and ensure the actual control performance,
the orders of the rational parts of the controller elements are determined as
the lowest such that a set of performance specifications on loop interaction
and loop performance are satisfied. Stability and robustness are analyzed and
substantial simulation shows that almost perfect control can be achieved.
 6.9 Notes and References 235

6.9 Notes and References

For plants with time delay, frequency domain methodologies are popular,
such as the inverse Nyquist array (Rosenbrock, 1974) and characteristic lo-
cus method (Macfarlane, 1980). Wang et al. (1997a) derived the decoupling
conditions for delay systems. The approach presented in this chapter is based
on Zhang (1999). In many cases, one replaces time delay by a rational func-
tion, say via Pade approximation, and then adopts delay-free designs (Morari
and Zafiriou, 1989). Such approximations will deteriorate when time delay
increases, and time delay compensation will be necessary to get high perfor-
mance and a topic of the next chapter.
7. Delay Systems

Time delay is a very common phenomenon encountered in process and chem-

ical industries. This is further complicated by multivariable nature of most
plants in operations. Input-output loops in a multivariable plant usually have
different time delays, and for a particular loop its output could be affected
by all the inputs through likely different time delays. As a result, such a
plant can be represented by a multivariable transfer function matrix having
multiple (or different) time delays around a operating point. The presence
of time delay in a feedback control loop could be a serious obstacle to good
process operation. It prevents high gain of a feedback controller from being
used, leading to sluggish system response. Such performance limitations can
be released only by time delay compensation.
For single variable processes with delay, The Smith control scheme can
remove the delay from the closed-loop characteristic equation and thus eases
feedback control design and improves set-point response greatly. An alter-
native scheme is the internal model control (IMC) structure. They can be
converted to each other. But the theory and design of the IMC is better de-
veloped than the Smith one. This chapter is to extend both the Smith and
IMC schemes to the multivariable case. An essential difference between the
SISO and MIMO with delay compensation is that for the SISO case the out-
put behavior is completely determined by the delay-free design shifted by the
plant time delay, whereas one can hardly tell the MIMO output performance
from the delay-free design because the delays in the process could mix up the
outputs of delay-free part to generate messy actual output response. Decou-
pling between the loops, in addition to serving its own purpose, can rectify
such a problem. Therefore, we will consider multivariable delay compensation
problem with decoupling in this chapter, and Sections 1 and 2 will address
the IMC and Smith schemes, respectively.

7.1 The IMC Scheme

The class of systems to be considered here is square, stable, nonsingular
MIMO linear ones. In process industry, most of the processes are open-loop
stable. A process usually has unequal numbers of inputs and outputs, but
normally a square subsystem will be selected from it for control purpose. A
Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 237-266, 2003.
Springer-Verlag Berlin Heidelberg 2003
238 7. Delay Systems

Fig. 7.1. Internal model control

singular process essentially cannot make independent control of each output

and should not be used.
The IMC control system is depicted in Figure 7.1, where G and Ĝ rep-
resent the transfer function matrices of the plant and its model, respectively,
and K the controller. Ĝ is assumed to be identical with G except our discus-
sion on robust stability:
 
g11 (s) · · · g1m (s)
 ..  ,
Ĝ(s) = G(s) =  ... ..
. .  (7.1)
gm1 (s) · · · gmm (s)

where

gij (s) = gij0 (s)e−Lij s ,

and gij0 (s) are strictly proper, stable, and scalar rational functions and Lij
are non-negative constants. It is also assumed that a proper input-output
pairing has been made for G(s) such that none of the m principal minors of
G(s) (the i th principle minor is the determinant of a (m−1)×(m−1) matrix
obtained by deleting the i th row and i th column of G(s)) is zero. Our task in
this section is to characterize all the realizable controllers K and the resultant
closed-loop transfer functions such that the IMC system is internally stable
and decoupled.
The closed-loop transfer matrix H between y and r can be derived from
Figure 7.1 as

H = GK[I + (G − Ĝ)K]−1 ,

which becomes

H = GK, if Ĝ = G.

Thus, the closed-loop is decoupled if and only if GK is decoupled (diago-

nal and nonsingular) and the IMC system is internally stable if and only if
 7.1 The IMC Scheme 239

K is stable. Therefore, the problem becomes to characterize all stable and

realizable controllers K and the resulting H such that GK = H is decoupled.
The first question to ask is whether or not a solution exists. For a stable
and nonsingular G, one may write it as

N (s)
G(s) = ,
d(s)

where d(s) is a least common denominator polynomial of G(s). Now choose

K(s) as

adj N (s)
K(s) = ,
σ(s)

where adj N (s) is the adjoint matrix of N (s) and σ(s) is a stable polynomial
and its degree is high enough to make the rational part of each kij proper.
Then, this K(s) is stable and realizable and it decouples G(s) since

|N (s)|
G(s)K(s) = Im ,
d(s)σ(s)

where | · | stands for the determinant of a matrix and Im is the m × m

identity matrix. Thus K(s) is a solution. Nonsingularity of G(s) is obviously
also necessary in order for GK to be decoupled.

Theorem 7.1.1. For a stable, square and multi-delay process G(s), the de-
coupling problem with stability via the IMC is solvable if and only if G(s) is
non-singular.

Let m = {1, 2, · · · , m} and Gij be the cofactor corresponding to gij in G.

It has been shown in Chapter 6 that in order for GK to be decoupled, we
have

GK = diag {g̃ii kii , i = 1, 2, · · · , m}, (7.2)

where
|G|
g̃ii = , ∀ i ∈ m; (7.3)
Gii
and

kji = ψji kii , ∀ i, j ∈ m, j = i, (7.4)

where
Gij
ψji = , ∀ i, j ∈ m, j = i. (7.5)
Gii
240 7. Delay Systems

One notes that for a given i, the resulting diagonal element of GK is indepen-
dent of the controller off-diagonal elements but contains only the controller
diagonal element kii .
To demonstrate how to find g̃ii and ψji , take the following as an example:
 1 e−2s −1 e−6s 
 s + 2 s +2 
G(s) =  2 . (7.6)
s − 0.5 e−3s (s − 0.5) e−8s
2 3
(s + 2) 2(s + 2)
Simple calculations give
2(s − 0.5)(s + 2)e−9s + (s − 0.5)2 e−10s
|G| = ,
2(s + 2)4

(s − 0.5)2 e−8s e−6s

G11 = , G21 = ,
2(s + 2)3 s+2

e−2s (s − 0.5)e−3s
G22 = , G12 = − .
s+2 (s + 2)2
It follows from (7.3) that the decoupled loops have their equivalent processes
as
2(s + 2)e−s + (s − 0.5)e−2s
g̃11 = ,
(s − 0.5)(s + 2)

2(s − 0.5)(s + 2)e−7s + (s − 0.5)2 e−8s

g̃22 = ,
2(s + 2)3
respectively. By (7.5), ψ21 and ψ12 are found as
2(s + 2) 5s
ψ21 = − e , ψ12 = e−4s .
s − 0.5
At the first glance, each diagonal controller kii may be designed for g̃ii
according to (7.2) and the off-diagonal controllers can then be determined
from (7.4). Recall that for SISO IMC design, a scalar process g(s) is first
factored into g(s) = g + (s)g − (s), where g + (s) contains all the time delay
and non-minimum phase zeros. The controller k(s) is then determined from
g(s)k(s) = g + (s)f (s), where f (s) is the IMC filter. This technique could
seemingly be applied to the decoupled equivalent processes g̃ii to design kii
directly, but it is actually not. The reasons are that in general g̃ii are not in
the usual format of a rational function plus time delay, and more essentially a
realizable kii derived from (7.2) based on g̃ii only may result in unrealizable or
unstable off-diagonal controllers kji , j = i, In other words, realizable kji , j =
i, may necessarily impose additional time delay and non-minimum phase
zeros to kii .
 7.1 The IMC Scheme 241

7.1.1 Analysis
In what follows, we will develop the characterizations of time delays and non-
minimum phase zeros on the decoupling controller and the resulting decou-
pled loops. For a nonsingular delay process G(s) in (7.1), a general expression
for g̃ii in (7.3) will be
M −αk s
k=0 nk (s)e
φ(s) = N , (7.7)
d0 (s) + l=1 dl (s)e−βl s
where nk (s) and dl (s) are all non-zero scalar polynomials of s, α0 < α1 <
· · · < αM and 0 < β1 < β2 < · · · < βN . Define the time delay for non-zero
φ(s) in (7.7) as

τ (φ(s)) = α0 .
It is easy to verify τ (φ1 φ2 ) = τ (φ1 ) + τ (φ2 ) and τ (φ−1 (s)) = −τ (φ(s)) for
any non-zero φ1 (s), φ2 (s) and φ(s). If τ (φ(s)) ≥ 0, it measures the time
required for φ(s) to have a non-zero output in response to a step input. If
τ (φ(s)) < 0 then the system output will depend on the future values of input.
It is obvious that for any realizable and non-zero φ(s), τ (φ(s)) can not be
negative. Therefore, a realizable K requires
τ (kji ) ≥ 0, ∀ i ∈ m, j ∈ Ji , (7.8)

where Ji
{j ∈ m | Gij = 0}. It follows from (7.4) that

Gij 
τ (kji ) = τ ( ii
kii ) = τ Gij ) − τ (Gii ) + τ (kii , ∀ i ∈ m, j ∈ Ji .
G

Thus, (7.8) is equivalent to

τ (kii ) ≥ τ (Gii ) − τ (Gij ), ∀ i ∈ m, j ∈ Ji ;
or
τ (kii ) ≥ τ (Gii ) − τi , ∀ i ∈ m, (7.9)
where
τi
min τ (Gij ). (7.10)
j∈Ji

Equation (7.9) is a characterization of the controller diagonal elements in

terms of their time delays, which indicates the minimum amount of time
delay that the i-th diagonal controller elements must contain. Consequently,
the resulting i-th diagonal elements of H = GK, hii = g̃ii kii , meet
τ (hii ) = τ (g̃ii ) + τ (kii ) ≥ τ (g̃ii ) + τ (Gii ) − τi , ∀ i ∈ m,
242 7. Delay Systems

which, by

|G|
τ (g̃ii ) = τ = τ (|G|) − τ (Gii ),
Gii

becomes

τ (hii ) ≥ τ (|G|) − τi , ∀ i ∈ m. (7.11)

Equation (7.11) is a characterization of the decoupled i − th loop transfer

function in terms of their time delays, which indicates the minimum amount
of time delay that the i − th decoupled loop transfer function must contain.
One concludes from the above development that for a stable and realizable
IMC controller K which decouples G, its diagonal elements are characterized
by (7.9) on their time delays and they will uniquely determine the off-diagonal
elements by (7.4). Equation (7.11) characterizes the time delays of the resul-
tant loops.
For illustration, consider again the example in (7.6). It follows from the
definition that the time delays for respective functions are τ (|G|) = 9,
τ (G11 ) = 8, τ (G12 ) = 3, τ (G21 ) = 6 and τ (G22 ) = 2. The values for τ1
and τ2 can be calculated from (7.10) as

τ1 = min τ (G11 ), τ (G12 ) = min{8, 3} = 3,

τ2 = min τ (G21 ), τ (G22 ) = min{6, 2} = 2.

It follows from (7.9) that k11 and k22 are characterized on their time delays
by

τ (k11 ) ≥ τ (G11 ) − τ1 = 8 − 3 = 5,

τ (k22 ) ≥ τ (G22 ) − τ2 = 2 − 2 = 0.

Also, h11 and h22 must meet

τ (h11 ) ≥ τ (|G|) − τ1 = 9 − 3 = 6,

τ (h22 ) ≥ τ (|G|) − τ2 = 9 − 2 = 7.

Consider now the performance limitation due to the non-minimum phase

zeros. Denote by C+ the closed right half of the complex plane (RHP). For
a non-zero transfer function φ(s), let Z+ φ be the set of all the RHP zeros
of φ(s), i.e., Z+
φ = {z ∈ C +
| φ(z) = 0}. Let ηz (φ) be an integer ν such
ν
that lims→z φ(s)/(s − z) exists and is non-zero. Thus, φ(s) has ηz (φ) zeros
at s = z if ηz (φ) > 0, or −ηz (φ) poles if ηz (φ) < 0, or neither poles nor
zeros if ηz (φ) = 0. It is easy to verify that ηz (φ1 φ2 ) = ηz (φ1 ) + ηz (φ2 ) and
 7.1 The IMC Scheme 243

ηz (φ−1 ) = −ηz (φ) for any non-zero transfer function φ1 (s), φ2 (s) and φ(s).
Obviously, a non-zero transfer function φ(s) is stable if and only if ηz (φ) ≥ 0,
∀ z ∈ C+ .
A stable K thus requires

ηz (kji ) ≥ 0, ∀ i ∈ m, j ∈ Ji , z ∈ C+ . (7.12)

It follows from (7.4) that

Gij
ηz (kji ) = ηz kii = ηz (Gij ) − ηz (Gii ) + ηz (kii ), ∀ i ∈ m, j ∈ Ji , z ∈ C+ .
Gii

Thus (7.12) is equivalent to

ηz (kii ) ≥ ηz (Gii ) − ηz (Gij ), ∀ i ∈ m, j ∈ Ji , z ∈ C+ ,

or

ηz (kii ) ≥ ηz (Gii ) − ηi (z), ∀ i ∈ m, z ∈ C+ , (7.13)

where

ηi (z)
min ηz (Gij ). (7.14)

j∈Ji

Since G(s) is stable, so are Gij . One sees that for ∀ i ∈ m and ∀ z ∈ C+ ,
ηi (z) ≥ 0. We have, for ∀ i ∈ m and ∀z ∈ C+ ,

ηz (Gii ) − ηi (z) ≤ ηz (Gii ),

and

ηz (Gii ) − ηi (z) = ηz (Gii ) − min ηz (Gij ) ≥ ηz (Gii ) − ηz (Gij )|j=i = 0.

j∈Ji

Thus, ηz (Gii ) − ηi (z) is bounded by

0 ≤ ηz (Gii ) − ηi (z) ≤ ηz (Gii ), ∀ i ∈ m, z ∈ C+ . (7.15)

Equation (7.15) also implies that kii need not have any non-minimum
phase zeros except at z ∈ Z+
Gii . Therefore, by (7.13), kii is characterized on
its non-minimum phase zeros by

ηz (kii ) ≥ ηz (Gii ) − ηi (z), ∀ i ∈ m, z ∈ Z+

Gii , (7.16)

with the right side bounded by (7.15).

Consequently, the resulting i-th diagonal elements of H = GK, hii =
g̃ii kii , meet
ηz (hii ) = ηz (g̃ii kii ) = ηz (g̃ii ) + ηz (kii ) ≥ ηz (g̃ii ) + ηz (Gii ) − ηi (z) ∀ i ∈ m, z ∈ C+ ,
244 7. Delay Systems

which, by

|G|
ηz (g̃ii ) = ηz = ηz (|G|) − ηz (Gii ),
Gii

becomes

ηz (hii ) ≥ ηz (|G|) − ηi (z), ∀ i ∈ m, z ∈ C+ . (7.17)

It is readily seen that for ∀ i ∈ m, z ∈ C+ ,

ηz (|G|) − ηi (z) ≤ ηz (|G|),

and
 
m

|G| 1
ηz (|G|) − ηi (z) = ηz = ηz  gij Gij 
(s − z)ηi (z) (s − z)ηi (z) j=1
 
m ij
G
= ηz  gij  ≥ 0.
j=1
(s − z)ηi (z)

Thus, ηz (|G|) − ηi (z) is bounded by

0 ≤ ηz (|G|) − ηi (z) ≤ ηz (|G|), ∀ i ∈ m, z ∈ C+ . (7.18)

Equation (7.18) also indicates that the i th closed-loop transfer function

hii need not have any non-minimum phase zeros except at z ∈ Z+ |G| . There-
fore, their characterization on non-minimum phase zeros are given by

ηz (hii ) ≥ ηz (|G|) − ηi (z), ∀ i ∈ m, z ∈ Z+

|G| , (7.19)

with the right side bounded by (7.18).

Consider the previous example again to demonstrate the method. It can
be found that Z+ + +
G11 = {0.5}, ZG22 = ∅, Z|G| = {0.5}, ηz (|G|)|z=0.5 = 1 and

ηz (G11 )|z=0.5 = 2, ηz (G12 )|z=0.5 = 1,

ηz (G21 )|z=0.5 = 0, ηz (G22 )|z=0.5 = 0.

By (7.14), one sees

η1 (z)|z=0.5 = min{ηz (G11 )|z=0.5 , ηz (G12 )|z=0.5 } = min{2, 1} = 1,

η2 (z)|z=0.5 = min{ηz (G21 )|z=0.5 , ηz (G22 )|z=0.5 } = min{0, 0} = 0.

Thus, k11 should then satisfy

 7.1 The IMC Scheme 245

ηz (k11 )|z=0.5 ≥ ηz (G11 )|z=0.5 − η1 (z)|z=0.5 = 2 − 1 = 1.

As Z+G22 = ∅, there will be no constraint for k22 on its non-minimum

phase zeros. By (7.19), the resultant loops meet

ηz (h11 )|z=0.5 ≥ ηz (|G|)|z=0.5 − η1 (z)|z=0.5 = 1 − 1 = 0,

ηz (h22 )|z=0.5 ≥ ηz (|G|)|z=0.5 − η2 (z)|z=0.5 = 1 − 0 = 1,

i.e., the second loop must contain a non-minimum phase zero at z = 0.5 of
multiplicity one while the first loop need not contain any non-minimum phase
zero.

Theorem 7.1.2. If the IMC system in Figure 7.1 is decoupled, stable and
realizable, then (i) the controller K’s diagonal elements are characterized by
(7.9) on its time delays and by (7.16) on its RHP zeros, and they uniquely
determine the controller’s off-diagonal elements by (6.20); (ii) the diagonal
elements of GK, hii = g̃ii kii , are characterized by (7.11) on its time delays
and (7.19) on its RHP zeros.

For the example in (7.6), choose the controller diagonal elements as

s − 0.5 −5s 1
k11 = e , k22 = ,
s+ρ s+ρ
where ρ > 0. It follows from the previous calculations that they satisfy the
conditions of Theorem 7.1.2 with lower bounds in (7.9) and (7.16) exactly
met. The off-diagonal controllers are then obtained from (7.4) and the com-
plete controller is
 
s − 0.5 e−5s 1 e−4s
 s + ρ s + ρ 
K= ,
−2(s + 2) 1
s+ρ s+ρ

which is both realizable and stable and the resulting GK is calculated as

 
2s + 4 + (s − 0.5)e−s −6s
 e 0 
(s + ρ)(s + 2)
GK =   −s
,

2s + 4 + (s − 0.5)e −7s
0 (s − 0.5) 3 e
2(s + ρ)(s + 2)

whose diagonal elements fall in our previous characterization on time delays

and non-minimum phase zeros, namely, the time delay for the decoupled
loops is no less than 6 and 7, respectively, and loop two must contain a
non-minimum phase zero at s = 0.5 of multiplicity 1.
However, if we choose the controller diagonal elements as
246 7. Delay Systems

1
k11 = k22 = ,
s+ρ
which violates the conditions of Theorem 7.1.2. It gives rise to
−2(s + 2)
k21 = e5s ,
(s + ρ)(s − 0.5)
which is neither realizable nor stable. The resultant h11 is given by
2s + 4 + (s − 0.5)e−s −s
h11 = g̃11 k11 = e ,
(s + ρ)(s + 2)(s − 0.5)
which is unstable, too. Such kii should be discarded.
A quite special phenomenon in such a decoupling control is that there
might be unstable pole-zero cancellations in forming g̃ii kii . In contrast to
the normal intuition, this will not cause instability. Notice from the previous
example that g̃11 has a pole at s = 0.5 and k11 has at least one zero at the
same location as ηz (k11 )|z=0.5 ≥ 1. Hence, an unstable pole-zero cancellation
at z = 0.5 occurs in forming g̃11 k11 . However, the resulting closed-loop system
is still stable because all the controller diagonal and off-diagonal elements are
stable. In fact, as far as G and K are concerned, there is no unstable pole-zero
cancellation since both G and K are stable.

7.1.2 Design
In this subsection, practical design of IMC is considered. We first briefly
review the IMC design for SISO case. Let g(s) be the transfer function
model of a stable SISO process and k(s) the controller. The resulting closed-
loop transfer function in case of no process-model mismatch is h(s), where
h(s) = g(s)k(s). It is quite clear that if h(s) has been specified then k(s)
is determined. However h(s) can not be specified arbitrarily. In IMC de-
sign, the process model g(s) is first factored into g(s) = g + (s)g − (s), where
g + (s) contains all the time delay and RHP zeros. Then h(s) is chosen as
h(s) = g + (s)f (s), where f (s) is the IMC filter, and consequently k(s) is
given by k(s) = h(s)g −1 (s).
The factorizations which yield g + (s) are not unique. Holt and Morari
(1985a,b) suggested that the following form be advantageous:
n
 zi − s
g + (s) = e−Ls , (7.20)
i=1
zi + s

where L is the time delay and z1 , z2 , · · · , zn are all the RHP zeros present
in g(s). The IMC filter f (s) is usually chosen in the form:
1
f (s) = ,
(τ s + 1)l
where τ and l are the filter time constant and order, respectively. The filter
time constant τ is a tuning parameter for performance and robustness.
 7.1 The IMC Scheme 247

For multivariable decoupling IMC control, the closed-loop transfer func-

tion matrix H(s) in case of no process-model mismatch is H = diag {hii } =
diag {g̃ii kii } and thus the key point in design lies in specifying the m de-
coupled loop transfer function hii (s). Unlike the SISO case, hii (s) can not
be chosen based solely on g̃ii (s) since this may cause an unrealizable and/or
unstable controller. Instead, hii (s) should be chosen according to Theorem
7.1.2. It is also obvious that for best performance of the IMC system and
simplicity of the controller it is undesirable to include any time delays and
non-minimum phase zeros in hii (s) more than necessary, that is, only the
minimum time delay in (7.11) and minimum multiplicity of unavoidable non-
minimum phase zeros in (7.19) should be taken. This leads to

 z−s
ηz (|G|)−ηi (z)
hii = e−(τ (|G|)−τi )s fi (s), ∀ i ∈ m, (7.21)
z+s
z∈Z+
|G|

where fi (s) is the i-th loop IMC filter. This choice of hii (s) will then determine
kii (s) as
−1
kii (s) = hii (s)g̃ii (s), (7.22)

which in turn yields off-diagonal elements kji (s) uniquely by (7.4).

Remark 7.1.1. Good control performance of the loops is what we seek apart
from the loop-decoupling requirement. However, we do not explicitly define
a performance measure and design the controller via the minimization (or
maximization) of some measure. These are done in model predictive control,
where the “size” of the error between the predicted output and a desired
trajectory is minimized; and in LQG control, where a linear quadratic index
is minimized to yield an optimal controller. These approaches have a dif-
ferent framework than ours. What we do here is to follow the IMC design
approach with the requirement of decoupling for the control of multivariable
multi-delay processes. Loop performance is simply observed with closed-loop
step responses for each output, in terms of traditional performance specifica-
tions such as overshoot, rise time and settling time in comparison with the
existing schemes of similar purposes. A link of the IMC design theory with
the quantitative error measure is that the item with the non-minimum phase
factor g + (s) in (7.20) used in IMC closed-loop transfer function in (7.21) can
actually minimize the H-2 norm of the error signal in the presence of a step
reference change (Morari and Zafiriou, 1989).

Remark 7.1.2. One may note that there is the case where good decoupling
and performance may not be both achievable and one is reached at cost of an-
other. For example, decoupling should never be used in flight control, instead,
couplings are deliberately employed to boost performance. This case is how-
ever not the topic of the present paper. We here address applications where
decoupling is required. Indeed, decoupling is usually required (Kong, 1995)
248 7. Delay Systems

in process control industry, for ease of process operations. Poor decoupling

in closed-loop system could, in many cases, lead to poor diagonal loop per-
formance. Thus, the design objective here is to decouple the loops at first
and then to achieve as good performance as possible for each loop. Loop per-
formance is considered to be good if best achievable performance is nearly
reached among all decoupled and stable loops, subject to limitations imposed
by delay, non-minimum phase zeros and bandwidth. Our design realizes this
by selecting the objective loop transfer functions in (7.21) in which each factor
reflects respective roles of delay, non-minimum phase zeros and bandwidth. It
might be worth noting that decoupling and loop performance may not always
be in conflict. It is shown (Linnemann and Wang, 1993) that for delay-free
stable systems some optimal controllers will automatically result in decoupled
loops though decoupling is not required in the problem formulation.
For a multivariable multi-delay process, the IMC controller designed by
the above procedure is usually in the complicated form of (7.7), which is
difficult to implement. Model reduction is then employed to find a much sim-
pler rational function plus dead time approximation to the above theoretical
controller. The topic of model reduction has been addressed in Chapter 2 in
a great detail and two effective algorithms given there. We can utilize one of
them for our present design.
Theoretically, if hii is specified as in (7.21) so that it contains no more
time delays than necessary, then each column of K has at least one element
whose time delay is zero. However, if the controller is obtained as a result of
the model reduction, that element might have non-zero time delay though it
is usually small. In such a case, it is reasonable to subtract it from all elements
in the column to speed up the loop response if it is positive or to make the
controller realizable if it is negative. This operation will not affect decoupling
and stability. The above development is summarized into the following design
procedure.

IMC design procedure. Seek a controller K(s) given G(s)

(i ) With i = 1 for loop one, apply model reduction to |G| and all non-
zero Gij . Take time delays and non-minimum phase zeros (including
multiplicities) from their reduced-order models. Find τi and ηi (z) for
all z ∈ Z+|G| from (7.10) and (7.14).
(ii ) Specify hii according to (7.21) and determine the diagonal controller kii
from (7.22). Apply model reduction to kii to obtain k̂ii .
(iii ) Calculate the controller off-diagonal elements kji from (7.4). Apply
model reduction to all kji to obtain k̂ji .
(iv ) Find the smallest time delay in k̂ji and subtract it from all the time
delays of k̂ji if it is non-zero.
(v ) Repeat the above steps for all other loops.
The IMC system in Figure 7.1 is referred to as the nominal case if G = Ĝ.
It is obvious that the IMC system is nominally stable if both the process G
 7.1 The IMC Scheme 249

and the controller K are stable, as discussed before. In the real world where
the model does not represent the process exactly, nominal stability is not
sufficient and robust stability of the IMC system has to be ensured. As a
standard assumption in robust analysis, we assume that the nominal IMC
system is stable (i.e., both Ĝ and K are stable). Suppose that the actual
process G is also stable and is described by
  
G∈ = {G : σ [G(jω) − Ĝ(jω)]Ĝ−1 (jω) = σ (∆m (jω)) ≤ ˜lm (ω)},

where ˆlm (ω) is the bound on the multiplicative uncertainty. It follows (Chap-
ter 3) that the IMC system is robust stable if and only if
−1
σ(Ĝ(jω)K(jω)) < lm (jω), ∀ω. (7.23)

Let K ∗ be the ideal controller obtained from (7.22) and (7.4), we have ĜK ∗ =
diag {hii }. Thus (7.23) becomes
 
σ diag {hii (jω)}K ∗ −1 (jω)K(jω) < lm −1
(jω), ∀ω.

One notes from (7.21) that diag {hii (jω)} = U diag {fi (jω)}, where U is a
unitary matrix. Noticing σ(U A) = σ(A) for a unitary U , thus a sufficient
and necessary robust stability condition is obtained as
 
σ diag {fi (jω)}K ∗ −1 (jω)K(jω) < lm −1
(jω), ∀ω. (7.24)

In practice, we can evaluate and plot the left hand side of (7.24) and compare
−1
with lm (jω) to see whether robust stability is satisfied given the multiplica-
tive uncertainty bound lm (jω). One also notes that if the error between the
ideal controller K ∗ and its reduced-order model K are small, i.e., K ∗ ≈ K,
then the robust stability condition becomes
−1
σ (diag {fi (jω)}) < lm (jω), ∀ω.

which indicates that the IMC filter should be chosen such that its largest
singular value is smaller than the inverse of the uncertainty magnitude for
all ω.

7.1.3 Simulation

In this subsection, several simulation examples are given to show the ef-
fectiveness of the decoupling IMC design. The effects of time delays and
non-minimum phase zeros are illustrated and control performance as well as
its robustness is compared with the existing multivariable Smith predictor
schemes. We intend to use commonly cited literature examples with real live
250 7. Delay Systems

background and try to explain technical points of the proposed method as

much as possible. The first and third examples are well-known distillation
column processes (Wood and Berry, 1973) and (Luyben, 1986a). The pro-
cess in the second example is a modification of the Wood and Berry process
where only the delays of its elements have been changed. This modification
generates non-minimum phase zeros for the multivariable process, which is
not available in the other two examples, and is used to illustrate relevant
design issues.

Example 7.1.1. Consider the well-known Wood/Berry binary distillation col-

umn plant (Wood and Berry, 1973):
 
12.8e−s −18.9e−3s
16.7s +1 21s + 1 
G(s) =  −7s
.
6.6e −19.4e−3s
10.9s + 1 14.4s + 1

1.5

y1

0.5

0
0 10 20 30 40 50 60 70 80 90 100

1.5
y2
1

0.5

−0.5
0 10 20 30 40 50 60 70 80 90 100

Fig. 7.2. Control Performance for Example 7.1.1

(—– IMC; - - - Jerome; · · · ideal IMC)
 7.1 The IMC Scheme 251

Step (i ) The application of model reduction to |G|, G11 and G12 produces
−0.1077s2 − 4.239s − 0.3881 −6.3s
|Ĝ| = e ,
s2 + 0.1031s + 0.0031
−1.349s − 42.81 −3s
Ĝ11 = e ,
s2 + 31.85s + 2.207
−0.6076s − 17.66 −7s
Ĝ12 = e .
s2 + 29.26s + 2.676
It is clear that τ (|Ĝ|) = 6.3, τ (Ĝ11 ) = 3, and τ (Ĝ12 ) = 7. Thus, by (7.10),
one finds τ1 = min{3, 7} = 3. Since |Ĝ| is of minimum phase, Z+ |Ĝ|
= ∅ and
there is no need to calculate η1 (z).
−3.3s
Step (ii ) According to (7.21), we have h11 = e s+1 with the filter chosen
1 −1
as s+1 . k11 is then obtained as k11 = g̃11 h11 from (7.22). The application of
model reduction to k11 yields
0.3168s2 + 0.0351s + 0.0012 −0.9s
k̂11 = e .
s2 + 0.1678s + 0.0074
Step (iii ) k21 is calculated from (7.4). Applying model reduction to k21
yields
0.1405s2 + 0.0180s + 0.0007 −4.9s
k̂21 = e .
s2 + 0.2013s + 0.0123
Step (iv ) The smallest time delay in k̂11 and k̂21 is 0.9 and then subtracted
from both elements’ time delays to generate the final k̂11 and k̂21 .
−5.3s
Step (v ) Loop two is treated similarly, where h22 = e s+1 . This results in
the overall controller as
 
0.3168s2 + 0.0351s + 0.0021 −0.2130s2 − 0.0222s − 0.0008 e−2s
 2
s + 0.1678s + 0.0074 2
s + 0.1399s + 0.0051 
K=
 2 2
.

0.1405s + 0.0180s + 0.0007 e−4s −0.1798s − 0.0206s − 0.0008
2 2
s + 0.2031s + 0.0123 s + 0.1607s + 0.0073

The set-point responses of the resulting IMC system (shortened as IMC)

are shown in Figure 7.2 with solid line. The corresponding set-point response
of the Jerome’s multivariable Smith predictor scheme with the controller
settings given in Jerome and Ray (1986) is depicted with dashed line in
the figure. The corresponding responses of the ideal IMC system H(s) =
diag {hii } are also shown with dotted line in the figure. The resulting ISE
(Integral Square Error) with equal weighting of the two outputs is calculated
as 7.16, 4.73 and 5.01 for the IMC design, Jerome’s scheme and the ideal
IMC control system, respectively. Although the ISE for the Jerome’s scheme
is smaller than that of the proposed method, the Jerome’s control system
is not robust and already exhibits very poor damping and severe oscillation
even for the nominal case, while the responses of the IMC are smooth and
the decoupling is almost perfect.
252 7. Delay Systems

1
10

0
10

−1
10

−1 0
10 10
w

Fig. 7.3. Robustness Analysis for Example 7.1.1

(—– left hand side of (7.24); - - - right hand side of (7.24))

y1 2

1.5

0.5

0
0 10 20 30 40 50 60 70 80 90 100

y2 2

1.5

0.5

−0.5
0 10 20 30 40 50 60 70 80 90 100

Fig. 7.4. Control System Robustness for Example 7.1.1

(—– IMC; - - - Jerome)
 7.1 The IMC Scheme 253

In order to see the robustness of the IMC design, we increase the dead
times of the process’ diagonal elements by 15%. In order to check the stability
of the perturbed system, the left hand side and right hand side of (7.24) are
plotted in Fig 7.3 with solid line and dashed line respectively. From the figure,
we conclude that the IMC system remains stable under the given perturba-
tions. The set-point responses are shown in Figure 7.4, exhibiting that the
performance of the IMC design is much superior compared with the Jerome’s
multivariable Smith predictor scheme. The resulting ISE is calculated as 7.76
for the IMC method while it is infinite for Jerome’s scheme due to instability.
♦

Example 7.1.2. Consider a variation of the above example:

 
g (s)e−s g120 (s)e−9s
G(s) = 110 ,
g210 (s)e−2s g220 (s)e−15s

where the delay free parts gij0 (s) are identical to Example 7.1.1. It follows
from the IMC procedure that

0.0374s2 + 3.72s − 0.5042 −13.6s

|Ĝ| = e ,
s2 + 0.1239s + 0.0041
−1.3493s − 42.81 −15s
Ĝ11 = e ,
s2 + 31.85s + 2.207
−0.6072s − 17.66 −2s
Ĝ12 = e .
s2 + 29.26s + 2.676

One readily sees that τ (|Ĝ|) = 13.6, τ (Ĝ11 ) = 15, τ (Ĝ12 ) = 2 and thus
τ1 = min{15, 2} = 2 from (7.10). A difference from Example 7.1.1 is
that |Ĝ| has a RHP zero at z = 0.136 with a multiplicity of one, i.e.,
Z+|Ĝ|
= {0.136} and ηz (|Ĝ|)|z=0.136 = 1. It is easy to find ηz (Ĝ11 )|z=0.136 = 0
and ηz (Ĝ12 )|z=0.136 = 0 since Ĝ11 and Ĝ12 are both of minimum phase.
We thus have η1 (z)|z=0.136 = min{0, 0} = 0 from (7.14) and h11 =
0.136−s −11.6s 1
(0.136+s)(s+1) e from (7.21) with the filter of s+1 . For the second loop,
1 0.136−s
with the filter chosen as s+1 , h22 is obtained as h22 = (0.136+s)(s+1) e−12.6s .
The rest of design is straightforward and gives
 
0.3656s2 + 0.0524s + 0.002 e−13.1s −0.2375s2 − 0.0294s − 0.0011 e−8s
 s + 0.2265s + 0.0124
2 2
s + 0.1790s + 0.0070 
K=
 2 2
.

0.1551s + 0.0291s + 0.012 −0.2003s − 0.0317s − 0.0014
s2 + 0.2726s + 0.0226 s2 + 0.2215s + 0.0136

For comparison, the Ogunnaike and Ray’s multivariable Smith predictor is

considered here, where the primary controller given in Ogunnaike and Ray
(1979) can be still used since the delay-free part of the process is identical for
our case and their case. The set-point responses shown in Figure 7.5 indicate
that the performance of the proposed IMC design is significantly better and
254 7. Delay Systems

y1
1.5

0.5

−0.5

−1
0 50 100 150 200 250 300

y2 1.5

0.5

−0.5

−1
0 50 100 150 200 250 300

Fig. 7.5. Control Performance for Example 7.1.2

(—– IMC; - - - Ogunnaike and Ray; · · · ideal IMC)

its response is very close to that of the ideal IMC controller (dotted line). The
resulting ISE is calculated as 52.5, 71.16 and 49.4 for the IMC, Ogunnaike
and Ray’s scheme, and the ideal IMC, respectively. ♦
Example 7.1.3. The process studied by Luyben (1986a) has the following
transfer function matrix:
 
0.126e−6s −0.101e−12s
60s + 1
G(s) =  (45s + 1)(48s + 1) .
0.094e−8s −0.12e−8s
38s + 1 35s + 1
It follows from the design procedure that with filters chosen as f1 (s) =
−6.5s −8.5s
1
f2 (s) = 3s+1 , h11 and h22 are obtained as h11 = e3s+1 and h22 = e3s+1 ,
respectively. The resultant controller is
 
146.8s2 + 6.355s + 0.0733 −6.027s2 − 0.3416s − 0.0053 e−9.3s
 2 2 
K =  s +20.3465s + 0.0034 s + 0.0395s + 0.0003
.
104.5s + 4.719s + 0.0561 −68.2s2 − 3.490s − 0.0527
s2 + 0.34s + 0.0034 s2 + 0.2496s + 0.0024

Since the multivariable Smith predictor design for this process is not avail-
able in the literature, the BLT tuning method (Luyben, 1986a) is used for
 7.1 The IMC Scheme 255

1.5
y1

0.5

0
0 100 200 300 400 500 600

1.5
y2
1

0.5

−0.5
0 100 200 300 400 500 600

Fig. 7.6. Control Performance for Example 7.1.3

(—– IMC; - - - BLT; · · · ideal IMC)

comparison. The set-point responses are shown in Figure 7.6, exhibiting that
significant performance improvement has been achieved with the IMC design
and again its response is very close to that of the ideal IMC system. The
resulting ISE is calculated as 17.7, 51.5 and 17.0 for the IMC, BLT, and the
ideal IMC system, respectively. ♦

In this section, an approach to the decoupling and stable IMC analysis

and design has been presented for multivariable processes with multiple time
delays. All the stabilizing controllers which solve this decoupling problem
and the resultant closed-loop systems are characterized in terms of their
unavoidable time delays and non-minimum phase zeros. Such delays and zeros
can be readily calculated from the process transfer function matrix and clearly
quantify performance limitations for any multivariable IMC system which is
decoupled and stable. It is interesting to note that a controller may necessarily
include some time delays and non-minimum phase zeros to make itself a
solution to the problem. A theoretical control design for the best achievable
performance is carried out based on the characterizations. Model reduction
is then exploited to simplify both analysis and design involved. Examples
have been given to illustrate our approach with which significant performance
improvement over the existing multivariable Smith predictor control methods
is evident.
256 7. Delay Systems

7.2 The Smith Scheme

In this section, a multivariable Smith predictor controller design is presented
for decoupling and stabilizing multivariable processes with multiple time de-
lays. A decoupler is first introduced in this new control scheme and it simpli-
fies the multivariable Smith predictor controller design to multiple single-loop
Smith predictor controller designs.

7.2.1 The Structure

Consider a multivariable process with the transfer matrix:

 
g11 (s) . . . g1m (s)
 
G(s) =  ... ..
. , (7.25)
gm1 (s) · · · gmm (s)

where

gij (s) = gij0 (s)e−Lij s ,

and gij0 (s) are strictly proper, stable scalar rational functions, and non-
negative Lij are the time delay associated with gij (s). Let the delay-free
part of the process be denoted by G0 = [gij0 ].

r + u y
 G
-
+
Ĝ
-
+ +
Ĝ0

Fig. 7.7. Multivariable Smith Predictor Control

The multivariable Smith predictor control scheme is shown in Figure 7.7

where G(s) and Ĝ(s) are the process and its model, respectively. Ĝ0 (s) is
the same as Ĝ(s) except that all the delays have been removed. C(s) is
the primary controller. When the model is perfect, i.e., Ĝ(s) = G(s) and
Ĝ0 (s) = G0 (s), the closed-loop transfer function from r to y becomes

H(s) = G(s)C(s)[I + G0 (s)C(s)]−1 .

 7.2 The Smith Scheme 257

It can be seen that I + G0 (s)C(s) contains no delays provided that C(s)

is so and it suggests that the primary controller C(s) can be designed with
respect to the delay free part G0 (s). This is the main attractiveness of the
scheme. However, unlike SISO case, even though C(s) is designed such that
H0 (s) = G0 (s)C(s)[I + G0 (s)C(s)]−1 has desired performance, the actual
system performance can not be guaranteed. This can be seen from the closed-
loop transfer function

H(s) = G(s)G−1
0 (s)G0 (s)C(s)[I + G0 (s)C(s)]
−1
= G(s)G0 (s)−1 H0 (s).

The actual system performance could be quite poor due to the existence
of G(s)G−10 (s). For the special case where the delays of all the elements in
each row of the transfer matrix are identical, the finite poles and zeros in
G(s)G−1 −1
0 (s) will all be cancelled. In this case, G(s)G0 (s) = diag e
−Lii s

and the system output is the delayed output of H0 (s). However, in general
this desired property is not preserved.
In order to overcome this problem and improve the performance of the
multivariable Smith predictor control system, we here propose a decoupling
Smith predictor control scheme depicted in Figure 7.8, where D(s) is a de-
coupler for G, Q(s) the decoupled process G(s)D(s), Q0 (s) is the same as
Q(s) except that all the delays are removed. Suppose that G(s)D(s) is de-
coupled, it is obvious that the Q(s) and Q0 (s) will be diagonal matrices. The
multivariable smith predictor design is then simplified to multiple single-loop
smith predictor designs for which various methods can be applied. This de-
coupling Smith scheme involves three parts, decoupler D(s), the decoupled
process Q and the primary controller C(s). Their design will be discussed in
the next subsection.

7.2.2 Design

For G(s)D(s) to be decoupled, it is shown in Chapter 6 that the elements of

the i − th column of D(s) should satisfy the conditions

|G|
GD = diag{ dii , i = 1, 2, · · · , m}, (7.26)
Gii

Gij
dji = dii := ψji dii , ∀i, j ∈ m, j = i. (7.27)
Gii
where m = {1, 2, · · · , m}. We may adopt the procedure in Section 1 to design
D. But the present case has a separate C(s) which will take care of all control
issues except decoupling which is the only duty of D(s). Thus, we wish to
get a simplest decoupling D(s) with minimum calculations. Let θji as the
smallest non-negative number such that ψji e−θji s does not have prediction
for all j ∈ m. Let θi = max θji . By choosing dii = e−θi s in (7.27), it is
j∈m,j=i
obvious that dji will have no predictions.
258 7. Delay Systems

Fig. 7.8. Decoupling Smith Control Scheme

Generally, for a multivariable system in (7.25), the dynamics of the resul-

tant dji , j = i, could be highly complicated. For ease of implementation, low-
order transfer functions are usually preferred. Yet, good performance should
be maintained. Thus, we apply model reduction to get a simpler, stable and
realizable approximation to each of them and use it for implementation.
Consider now Q in Figure 7.8. In the case of perfect decoupling, Q should
be
|G|
Q = GD = diag{ ii dii , i = 1, 2, · · · , m}
G
= diag{q11 , q22 , · · · , qmm }.
Generally, qii is too complicated to implement. Thus, model reduction is also
applied to qii to obtain a simpler yet good approximation for implementation.
And Q0 in Figure 7.8 is readily obtained as the delay free part of Q(s).
With decoupling by D, the multivariable smith predictor design for Figure
7.8 is now simplified to multiple single-loop smith predictor control designs.
Let the primary controller be

C(s) = diag{ c11 (s) · · · cmm (s)} .

Each individual cii (s) is designed with respect to the delay free part qii0 of
qii such that closed-loop system formed by cii (s) and qii0 has the desired
performance. This is a number of SISO problems and deserves no further
discussions.

7.2.3 Stability Analysis

The multivariable Smith system in Figure 7.8 is referred to as the nominal

case if GD = Q. As the process G in (7.25) is assumed to be stable and the
decoupler D is designed to be stable, the control system in Figure 7.8 will
be stable if and only if the primary controller C stabilizes Q0 .
The nominal condition GD = Q can be violated in practice for several rea-
sons. Firstly, due to the approximation of the decoupler by model reduction,
 7.2 The Smith Scheme 259

Fig. 7.9. Uncertain Smith System

the transfer matrix GD may not be diagonal whereas Q is implemented as

diagonal form. Secondly, approximation of qii by model reduction also causes
discrepancy between GD and Q even if the decoupling is perfect. Most im-
portantly, in the real world, the model G(s) may not represent the actual
process exactly. For robustness analysis, the actual process is assumed to be
any member of a family of possible processes. It is shown by an example
in Palmor (1980) that the Smith predictor controller could be unstable for
infinitesimal perturbation in the dead time, even though it may be nominally
stable. Therefore, nominal stability is not sufficient. Robust stability of the
closed-loop has to be ensured for the application of the multivariable Smith
system.
Let the nominal stable process transfer function be Ĝ(s) and the real
process be described by the family:
 
Π = G(s) : G(s) = Ĝ(s) + ∆(s), σ̄(∆(jω)) ≤ |γ(jω)| , (7.28)

where the perturbation ∆(s) is stable. The multivariable Smith predictor

control system in Figure 7.8 is said to be robustly stable if the closed-loop
system is stable for each member in the family Π. This uncertain Smith
system is shown in Figure 7.9, and this can be redrawn into the standard
form (Chapter 3), with ∆(s) appearing in a feedback path, as in Figure 7.10.
Note that if we assume that the Smith system is nominally stable, then the
dashed line encloses a stable system. Let F (s) be the transfer function from
z̃ to ũ in Figure 7.10. It follows that the uncertain Smith system is robustly
stable, if and only if
σ̄[F (jω)]|γ(jω)| < 1, ω ∈ [0, ∞), (7.29)
or equivalently
F (s)γ(s) ∞ < 1, (7.30)
where σ̄(·) denotes the largest singular value of a matrix, and · ∞ the H∞ -
norm of a transfer matrix. To find F , by Figure 7.10, we have
260 7. Delay Systems

Fig. 7.10. Equivalent Uncertain Smith System

ũ = DCe, (7.31)

e = r − y + (Q − Q0 )Ce, (7.32)

y = ĜDCe + z̃. (7.33)

Collecting (7.31) to (7.33) yields

ũ = DC[I + ĜDC − (Q − Q0 )C]−1 r − DC[I + ĜDC − (Q − Q0 )C]−1 z̃.

The transfer function from z̃ to ũ is then given by

F (s) = −DC[I + ĜDC − (Q − Q0 )C]−1 .

Theorem 7.2.1. Assume that the family of stable processes Π is described

by (7.28) and that the nominal closed-loop system is stable. Then the mul-
tivariable Smith predictor control system in Figure 7.9 is robustly stable if
and only if (7.29) or (7.30) holds true.

7.2.4 Simulation

Several simulation examples are now given to show the effectiveness of the
decoupling Smith control.
 7.2 The Smith Scheme 261

Example 7.2.1. Consider the well-known Wood/Berry binary distillation col-

umn process (Wood and Berry, 1973):
 −s −3s

12.8e −18.9e
16.7s+1 21s+1
G(s) =  .
6.6e−7s −19.4e−3s
10.9s+1 14.4s+1

The diagonal elements of decoupler D(s) are chosen as d11 (s) = 1 and
d22 (s) = 1 since the ψ12 = g12 /g11 and ψ21 = g21 /g22 in (7.27) have no pre-
dictions. Model reduction yields the off-diagonal elements of the decoupler
as
(24.66s + 1.48)e−2s
d12 = ,
21s + 1
and
(4.90s + 0.34)e−4s
d21 = ,
10.9s + 1
and the decoupler is formed as
 
(24.66s+1.48)e−2s
1 21s+1
D(s) =  .
(4.90s+0.34)e−4s
10.9s+1 1

The application of the model reduction to gii (s), the diagonal elements of
G(s)D(s), produces

6.37e−0.92s
g11 = ,
0.61s2 + 5.45s + 1
and
−9.65e−3.31s
g22 = ,
4.59s + 1
giving
 
6.37e−0.92s
0.61s2 +5.45s+1 0
Q(s) =  
−9.65e−3.31s
0 4.59s+1 ,

and
 6.37

0.61s2 +5.45s+1 0
Q0 (s) = −9.65
.
0 4.59s+1

respectively, The primary controller is designed as simple PID:

262 7. Delay Systems

Fig. 7.11. Control Performance for Example 7.2.1 (— Proposed; - - - Jerome)

 0.066

0.23 + s − 0.09s 0
C(s) = 0.048
.
0 −0.14 − s + 0.097s

The set-point response is shown in Figure 7.11. For comparison, the Jerome’s
multivariable Smith predictor scheme is used with the controller settings
from Jerome and Ray (1986). It can be seen that the responses of the de-
coupling Smith control are much better and the decoupling is almost per-
fect. ♦
Example 7.2.2. Wardle and Wood (1969) give the following transfer function
matrix model for an industrial distillation column:
 −6s −12s

0.126e −0.101e
60s+1 (45s+1)(48s+1)
G(s) =  .
0.094e−8s −0.12e−8s
38s+1 35s+1

For this process, the diagonal elements of the decoupler D(s) are chosen as
d11 (s) = 1 and d22 (s) = 1 since the ψ12 = g12 /g11 and ψ21 = g21 /g22 have no
predictions. Model reduction yields

 
(48.09s+0.802)e−6s
1 2160s2 +93s+1 
D(s) =  .
47.72s2 +28.78s+0.783
66.14s2 +39.74s+1 1
 7.2 The Smith Scheme 263

Fig. 7.12. Control Performance for Example 7.2.2 (— Proposed; - - - Jerome)

The reduced-order modelling of the diagonal elements of G(s)D(s) yields

 
(19.54s+1)e−6.52s
8827.65s2 +712.54s+21.33 0
Q(s) =  ,
(57.16s+1)e−8.6s
0 15578s2 +964.2s+22.4

and Q0 (s) is readily obtained by taking the delay free part of Q(s). The
primary controller is then designed as
 
205.4 + 11.1
s − 47.75s 0
C(s) = .
0 −73.0 − 7.02
s − 17.2s

For comparison, the BLT tuning method (Luyben, 1986) is used here. The
set-point responses are shown in Figure 7.12, exhibiting that significant
performance improvement has been achieved with the decoupling Smith
design. ♦

Example 7.2.3. Consider the following system in Jerome and Ray (1986)
 −2s −4s 
(1−s)e 0.5(1−s)e
s2 +1.5s+1 (2s+1)(3s+1)
G(s) =  .
0.33(1−s)e−6s (1−s)e−3s
(4s+1)(5s+1) 4s2 +6s+1

The design procedure yields

264 7. Delay Systems

Fig. 7.13. Control Performance for Example 7.2.3

 2

+0.75s+0.5)e−2s
1 − (0.5s 6s2 +5s+1
D(s) =  2
.
+1.98s+0.33)e−3s
− (1.32s 20s2 +9s+1 1

 
(16.9s2 +10.7+1)e−4.54s
17.4s2 +9.77s+1.2 0
Q(s) =  ,
(5.18s+1)e−4.83s
0 26.79s2 +9.72s+1.2

and
 0.37

−0.22 + s + 0.11s 0
C(s) = 0.16
.
0 1.79 + s − 0.90s

The set-point change response is shown in Figure 7.13. For compari-

son the Jerome’s multivariable Smith predictor (Jerome and Ray, 1986) is
used, where an augmented compensator is required in their controller. The
results indicate that the decoupling Smith method also gives improvement in
performances.
In order to see the robustness of the decoupling Smith method, we in-
crease the static gains and dead-times of the process’ diagonal elements
by 20% and 30%, respectively. The set-point response is depicted in Figure
7.14, showing that the performance of the decoupling Smith scheme is much
superior compared with the Jerome’s multivariable Smith predictor con-
troller. ♦
 7.3 Notes and References 265

Fig. 7.14. Control System Robustness for Example 7.2.3 (— Proposed; - - - Jerome)

In this section, a stable and decoupling multivariable Smith predictor

scheme is presented for multivariable processes with multiple time delay. Our
approach to the decoupler design leads to a number of independent single-
loop Smith predictor designs. Both nominal and robust stability analysis
is given. Examples have been given to illustrate our approach with which
significant performance improvement over the existing multivariable Smith
predictor control methods is observed. This design is simpler, compared with
the IMC control in Section 1.

7.3 Notes and References

The first time delay compensation scheme was proposed by Smith (1957) for
SISO plants and is named after him. This scheme was extended to multivari-
able systems with single delay by Alevisakis and Seborg (1973),Alevisakis and
Seborg (1974) and with multi-delays by Ogunnaike and Ray (1979),Ogun-
naike et al. (1983). Due to multivariable interactions, the performance of such
MIMO Smith schemes might sometimes be poor ((Garcia and Morari, 1985).
Jerome and Ray (1986) proposed a different version of multivariable Smith
predictor control to improve overall performance. However, their design of
the primary controller is based on a transfer matrix with time delays and is
difficult to carry out. This also contradicts the Smith’s philosophy of delay
removal from closed-loop characteristic equation for ease of stable design. The
266 7. Delay Systems

work presented in Section 2 is based on Wang et al. (2000) and can achieve
better performance.
The IMC was introduced by Garcia and Morari (1982) and studied thor-
oughly in Morari and Zafiriou (1989) as a powerful control design strategy for
linear systems. In principle, if a multivariable delay model, Ĝ(s), is factor-
ized as Ĝ(s) = Ĝ+ (s)Ĝ− (s) such that Ĝ+ (s) is diagonal and contains all the
time delays and non-minimum phase zeros of Ĝ(s), then the IMC controller
can be designed, like the scalar case, as K(s) = {Ĝ− (s)}−1 with a possible
filter appended to it, and the performance improvement of the resultant IMC
controller over various versions of the multivariable Smith predictor schemes
can be expected (Garcia and Morari, 1985). For a scalar transfer function, it
is quite easy to obtain Ĝ+ (s). However, it becomes much more difficult for a
transfer matrix with multi-delays. The factorization is affected not only by
the time delays in individual elements but also by their distributions within
the transfer function matrix, and a non-minimum phase zero is not related to
that of elements of the transfer matrix at all (Holt and Morari, 1985b; Holt
and Morari, 1985a). The systematic method in Section 1 can effectively deal
with such a problem and it is from Zhang (1999).
8. Near-Decoupling

Several approaches to decoupling control have been presented in details in the

previous chapters and plant uncertainties not addressed explicitly in design
though robust stability is checked sometimes. It is noted that exact decou-
pling may not be necessary even in the nominal case for many applications
and usually impossible in the uncertain case. What is really needed instead is
to limit the loop interaction to an acceptable level. In this chapter, we propose
the concept of near-decoupling and develop an approach to near-decoupling
controller design for both nominal and uncertain systems.
This chapter is organized as follows. In Section 8.1, we provide an example
to illustrate the necessity for proposing the new concept of near-decoupling.
Its exact definition is then given in Section 8.2. Fundamental lemmas for
solving near-decoupling control problem and the basic result on the near-
decoupling controller design for the case of exact model and state feedback
are also presented in this section. The design method is then extended to
the cases of (i) uncertain models and state feedback, (ii) exact models and
dynamic output feedback, and (iii) uncertain models and dynamic output
feedback respectively in Sections 8.3, 8.4 and 8.5. A numerical illustrative
example is presented in Section 8.6 to show the whole design procedure of
our approach. Finally concluding remarks are drawn in Section 8.7.
Notations: Rn denotes n−dimensional Euclidean space, Rn×m is the set
of all n × m real matrices, I is an identity matrix whose dimension is implied
from context, k · k2 refers to either the Euclidean vector norm or the induced
matrix 2-norm, and kHk∞ represents the ∞-norm of a transfer matrix H(s)
in H∞ space. σ(M ) and λmin (M ) denote the smallest singular value and
minimum eigenvalue of a matrix M , respectively. M ∗ stands for the complex
conjugate transpose of M . The notation X > Y (or X ≥ Y, respectively)
means that X − Y is positive definite (or positive semidefinite, respectively).

8.1 Motivation for the Concept of Near-Decoupling

Consider the following 2 × 2 system

ẋ = A0 x + B0 u, y = C0 x,

where
Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 267-292, 2003.
Springer-Verlag Berlin Heidelberg 2003
268 8. Near-Decoupling

   
−1 0 0 1 0  
100
A0 =  0 −2 0  , B0 =  2 3  , C0 = .
111
0 0 −3 −3 −3

Its transfer matrix is

" 1 #
s+1 0
H0 (s) = 4s+6 3
.
(s+1)(s+2)(s+3) (s+2)(s+3)

It is found in Chapter 4 that the above system can be decoupled by the

following state feedback

u = Kx + Gv,

where
" # " #
1 0 0 1 0
K= , G= .
− 53 − 43 −3 − 43 1
3

Indeed, the resulting closed-loop transfer matrix is

"1 #
s 0
H0cl (s) = .
0 s12

Now suppose that there are some parameter perturbations in matrix A0 , i.e.,
suppose
 
−1 + a1 0 0
A0 =  0 −2 + a2 0 ,
0 0 −3 + a3

Under the same decoupling law, the closed-loop transfer matrix of the per-
turbed system is as follows
" 1 #
s−a1 0
Hcl (s) = (a3 −2a2 +a1 )s+a1 a2 +a2 a3 −2a1 a3 −a1 +4a2 −3a3 a −a +1
.
2 3
(s−a1 )(s2 −(a2 +a3 )s+6a2 −6a3 +a2 a3 ) s2 −(a2 +a3 )s+6a2 −6a3 +a2 a3

Thus it is clear that the decoupling still holds only when


a3 − 2a2 + a1 = 0,
a1 a2 + a2 a3 − 2a1 a3 − a1 + 4a2 − 3a3 = 0.

This is the case if


a1 = a3 ,
a2 = a3 ,
or
 8.2 Near-Decoupling for Unperturbed Systems: State Feedback 269


a1 = a3 − 2,
(8.1)
a2 = a3 − 1.

The relationship (8.1) is inadmissable since it leads to Hcl (2, 2) ≡ 0. Therefore

only when a1 = a2 = a3 , does the perturbed system remain decoupled. If we
regard a := (a1 , a2 , a3 ) as a point in R3 and a1 , a2 and a3 vary independently
in the space R3 , then the exact decoupling can be achieved only when a
hits the line {(a1 , a2 , a3 ) ∈ R3 : a1 = a2 = a3 }. It is well known that the
probability that a point randomly distributed in the space R3 hits a line is
zero in the sense of geometric probability measure.
It is clear that the above procedure is applicable to general multiinput
multioutput (MIMO) linear systems with parameter perturbations and it
can be shown that a (fixed) decoupling law can decouple a perturbed MIMO
system only when the parameter perturbations are in a subspace of the space
consisting of all free parameter perturbations. Therefore we can generally say
that exact decoupling cannot be achieved almost for every perturbed system.
Based upon the aforementioned fact, we will propose a new concept for
decoupling: near-decoupling. According to this concept, our aim of designing
an near-decoupling law is to make the magnitude of off-diagonal elements in
the transfer matrix of a system as small as possible, instead of zero, under
the existence of parameter perturbations, while to maintain the magnitude
of diagonal elements in the transfer matrix to be larger than a given number.

8.2 Near-Decoupling for Unperturbed Systems: State

Feedback

In this section, we first define the concept of near-decoupling and then develop
a method to find solutions for near-decoupling controllers using state feedback
for systems with no perturbations, which will lay a basis for the developments
of the sections to be followed. Let us consider a linear time-invariant system
(A, B, C, D) described by

ẋ = Ax + Bu,
(8.2)
y = Cx + Du,

where x ∈ Rn denotes the state, u ∈ Rm denotes the control input, y ∈ Rl

denotes the output, and A, B, C and D are constant matrices with appro-
priate dimensions. Pp The output y is partitioned in p block outputs yi , each
of size li , with i=1 li = l. Similarly, P The input u is also partitioned in p
p
block inputs ui , each of size mi , with i=1 mi = m. The block subsystems
(A, Bj , Ci , Dij ) of (A, B, C, D) are introduced where Ci , Bj and Dij are the
ith block of C, jth block of B and ij-th block of D, respectively, in connec-
tion with the splitting of y = [y1T , y2T , · · · , ypT ]T and u = [uT1 , uT2 , · · · , uTp ]T .
Thus for such a splitting, we write
270 8. Near-Decoupling

   
C1 D11 D12 · · · D1p
 C2   D21 D22 · · · D2p 
   
B = [B1 , B2 , · · · , Bp ], C =  . , D= . .. .. . .
 ..   .. . . .. 
Cp Dp1 Dp2 · · · Dpp

For the convenience of later exposition, we write

 
Di = Di1 Di2 · · · Dip .

Correspondingly, the transfer matrix of system (8.2) is also split as

 
H11 (s) H12 (s) · · · H1p (s)
 H21 (s) H22 (s) · · · H2p (s) 
 
H(s) = C(sI − A)−1 B + D =  . .. .. .. 
 .. . . . 
Hp1 (s) Hp2 (s) · · · Hpp (s)

with

Hij (s) = Ci (sI − A)−1 Bj + Dij , i, j = 1, 2, . . . , p.

Definition 8.2.1. System (8.2) is said to be near-decoupled for the input-

output pairs (ui , yi ), i = 1, 2, . . . , p, if
(i) system (8.2) is stable;
(ii) the following inequalities
σ(Hii (jω)) ≥ βi , ∀i ∈ {1, 2, . . . , p}, ∀ω ∈ (−∞, +∞), (8.3)
kHij k∞ < γ, ∀i, j ∈ {1, 2, . . . , p}, i 6= j. (8.4)
hold for given numbers βi > 0, i = 1, 2, . . . , p, and γ > 0.

Definition 8.2.2. System (8.2) is said to be nearly state-feedback decou-

plable if there exists a state feedback law

u = Kx + v (8.5)

such that the closed loop system (8.2)-(8.5) is near-decoupled for the new
input-output pairs (vi , yi ), i = 1, 2, . . . , p, in the sense of Definition 8.2.1.

Notice that the parameter γ in the above definitions should be sufficiently

small, whose value can be either specified by the designer or minimized in
the controller design process. In the first case, we can generally specify, for
example, γ = 10−3 × min{β1 , β2 , . . . , βp }.
To make the design of controller (8.5) be numerically tractable, it is highly
desirable to relate the performance indices described in Definition 8.2.1 di-
rectly to the system parameters (A, B, C, D). For requirement (8.4), it can
be dealt with by the following lemma, which has become a standard result
in H∞ control.
 8.2 Near-Decoupling for Unperturbed Systems: State Feedback 271

Lemma 8.2.1. (Boyd et al., 1994) Denote by H(s) the transfer matrix of a
system (A, B, C, D). Then the condition

kHk∞ < γ

is equivalent to ∃P > 0 such that

 T 
A P + P A P B CT
 BT P −γI DT  < 0.
C D −γI

To deal with the requirement (8.3), let us first establish the following
result.
Lemma 8.2.2. Consider system (8.2) with transfer matrix H(s). Then the
inequality

σ(H(jω)) ≥ β, ∀ω ∈ (−∞, +∞), (8.6)

holds if and only if

Z +∞ Z +∞
T 2
y (t)y(t)dt ≥ β uT (t)u(t)dt, ∀u ∈ L2 (−∞, +∞) (8.7)
−∞ −∞

holds.
Proof. The only if part. First notice, in terms of Parseval’s theorem, the fact
that inequality (8.7) is equivalent to the following inequality:
Z +∞ Z +∞
∗ 2
Y (jω)Y (jω)dω ≥ β U ∗ (jω)U (jω)dω ∀U ∈ H2 , (8.8)
−∞ −∞

where U (jω) and Y (jω) denote the Fourier’s transform of u(t) and y(t) re-
spectively. For a brief introduction on the spaces L2 (−∞, +∞) and H2 , read-
ers is referred to (Zhou et al., 1996). Inequality (8.6) means that

H ∗ (jω)H(jω) ≥ β 2 I, ∀ω ∈ (−∞, +∞).

Thus we have
Z +∞ Z +∞
1
y T (t)y(t)dt = Y ∗ (jω)Y (jω)dω
−∞ 2π −∞
Z +∞
1
= U ∗ (jω)H ∗ (jω)H(jω)U (jω)dω
2π −∞
Z
β 2 +∞ ∗
≥ U (jω)U (jω)dω
2π −∞
Z +∞
= β2 uT (t)u(t)dt.
−∞
272 8. Near-Decoupling

Now let us prove the if part. Suppose, by contradiction, that inequality

(8.6) does not hold for all ω ∈ (−∞, +∞). Without loss of generality, we tem-
porarily assume m ≤ l. According to singular value decomposition theorem,
there exist unitary matrices EL (jω) and ER (jω) such that
∗
H(jω) = EL (jω)Λ(jω)ER (jω)

where
 
σ1 (jω) 0 ··· 0
 0 σ2 (jω) · · · 0 
 
 .. .
.. . .. .
.. 
Λ(jω) =  . 
 
 0 0 · · · σm (jω) 
0 0 ··· 0

Notice that the last row of zeros actually includes l − m rows of zeros. Now
we have
 
|σ1 (jω)|2 0 ··· 0
 0 |σ (jω)|2
· ·· 0 
 2  ∗
H ∗ (jω)H(jω) = ER (jω)  .. .. . ..  ER (jω)
 . . .. . 
0 0 · · · |σm (jω)|2

Suppose that there exists an ω0 such that

|σm (jω0 )|2 < β 2 .

Since σm (jω) is continuous with respect to ω, it follows that there exists a

closed interval I with ω0 ∈ I and |I| 6= 0, where |I| denotes the length of
the set I, such that

|σm (jω)|2 < β 2 , ∀ω ∈ I.

Notice that there might be many such closed intervals and the length of
the such interval might be infinite. In this case, we just choose one of such
intervals and limit its length to be finite. Define em = [0 · · · 0 1]T ∈ Rm .
Now we choose the signal U (jω) as follows

ER (jω)em when ω ∈ I,
U (jω) =
0 otherwise.

Such a choice on U (jω) will lead to

Z +∞
U ∗ (jω)U (jω)dω = |I| (8.9)
−∞

and
 8.2 Near-Decoupling for Unperturbed Systems: State Feedback 273

Z +∞
Y ∗ (jω)Y (jω)dω
−∞
Z
= ∗
eTm ER (jω)ER (jω)Λ∗ (jω)Λ(jω)ER
∗
(jω)ER (jω)em dω
I
Z
= |σm (jω)|2 dω
I

< β 2 |I|. (8.10)

Combining (8.9) and (8.10) gives
Z +∞ Z +∞
Y ∗ (jω)Y (jω)dω < β 2 U ∗ (jω)U (jω)dω,
−∞ −∞

which contradicts the inequality (8.7). Thus the proof is completed. 

Lemma 8.2.3. Denote by H(s) the transfer matrix of system (8.2). If there
exists a matrix P > 0 such that
 
−AT P − P A + C T C −P B + C T D
≥ 0, (8.11)
−B T P + DT C −β 2 I + DT D

the following inequality

σ(H(jω)) ≥ β, ∀ω ∈ [−∞, +∞], (8.12)

holds.

Proof. Let V (x) = −xT P x. Consider the motion of system (8.2) with zero
initial condition, i.e., x(0) = 0. If one can prove that

dV (x)
+ y T (t)y(t) − β 2 uT (t)u(t) ≥ 0, ∀u ∈ L2 (−∞, +∞), (8.13)
dt
then one has
Z +∞
[y T (t)y(t) − β 2 uT (t)u(t)]dt
0
Z +∞
dV (x)
≥− dt = xT (+∞)P x(+∞) − xT (0)P x(0) ≥ 0.
0 dt
Thus
Z +∞ Z +∞
T 2
y (t)y(t)dt ≥ β uT (t)u(t)dt, ∀u ∈ L2 (−∞, +∞).
0 0

From Lemma 8.2.2, we can therefore conclude that inequality (8.12) holds
true.
274 8. Near-Decoupling

Now simple algebra yields that (8.13) means

   T  
T T −(AT P + P A) −P B C x
[x u ] + [C D] ≥ 0.
−B T P −β 2 I DT u

Thus it is clear that if inequality (8.11) holds, inequalities (8.13) and then
(8.12) hold too. 

Corollary 8.2.1. For system (8.2), if there exists a matrix P > 0 such that
 
−AT P − P A −P B + C T D
≥ 0,
−B T P + DT C −β 2 I + DT D

then inequality (8.12) holds.

Remark 8.2.1. From Lemma 8.2.3 it can be seen that the parameter β is
upper bounded by
q
β ≤ λmin (DT D).

Based upon the above development, we can establish the following theo-
rem.

Theorem 8.2.1. For system (8.2), if there are matrices Q > 0 and F such
that the following LMIs
 
−QAT − AQ − F T B T − BF −Bi + (QCiT + F T DiT )Dii
> 0,
−BiT + DiiT
(Ci Q + Di F ) −βi2 I + Dii
T
Dii
∀i ∈ {1, 2, . . . , m}; (8.14)
 
QAT + AQ + F T B T + BF Bj QCiT + F T DiT
 BjT −γI T
Dij  < 0,
Ci Q + Di F Dij −γI
∀i, j ∈ {1, 2, . . . , m}, i 6= j. (8.15)
hold, then system (8.2) is nearly state-feedback decouplable. The state-feedback
law is given by

u = Kx + v, K = F Q−1 (8.16)

and the closed-loop performance indices satisfy

σ(Hcl,ii (jω)) ≥ βi , ∀i ∈ {1, 2, . . . , p}, ∀ω ∈ [−∞, +∞], (8.17)
kHcl,ij k∞ < γ, ∀i, j ∈ {1, 2, . . . , p}, i 6= j, (8.18)
p
where βi < λmin (Dii T D ), i = 1, 2, . . . , p.
ii
 8.2 Near-Decoupling for Unperturbed Systems: State Feedback 275

Proof. System (8.2) under the control (8.16) reads as

ẋ = (A + BK)x + Bv,
y = (C + DK)x + Dv.
Applying Lemma 8.2.1 and Corollary 8.2.1 to the block transfer matrix
Hcl,ij (s) = trans(A + BK, Bj , (C + DK)i , Dij ) = trans(A + BK, Bj , Ci +
Di K, Dij ), i, j ∈ {1, 2, . . . , p} (where trans refers the transfer matrix of a
state-space model) yields that performance requirements (8.17)-(8.18) are
satisfied if there exists a matrix P > 0 such that the following matrix in-
equalities
 
−(A + BK)T P − P (A + BK) −P Bi + (Ci + Di K)T Dii
> 0,
−BiT P + Dii T
(Ci + Di K) −βi2 I + Dii
T
Dii
∀i ∈ {1, 2, . . . , m}; (8.19)
 
(A + BK)T P + P (A + BK) P Bj (Ci + Di K)T
 BjT P −γI T
Dij  < 0,
(Ci + Di K) Dij −γI
∀i, j ∈ {1, 2, . . . , m}, i 6= j (8.20)
−1
hold. Let Q = P and F = KQ. Then we have
   
Q0 −(A + BK)T P − P (A + BK) −P Bi + (Ci + Di K)T Dii Q0
0 I −BjT P + Dii T
(Ci + Di K) −β 2 I + Dii
T
Dii 0 I
 T T T T T T

−QA − AQ − F B − BF −Bi + (QCi + F Di )Dii
= ,
−BiT + Dii T
(Ci Q + Di F ) −βi2 I + Dii
T
Dii
   
Q00 (A + BK)T P + P (A + BK) P Bj (Ci + Di K)T Q00
 0 I 0 BjT P −γI T
Dij  0 I 0
0 0I (Ci + Di K) Dij −γI 0 0I
 T T T T T T

QA + AQ + F B + BF Bj QCi + F Di
= BjT −γI DijT ,
Ci Q + Di F Dij −γI

i.e., matrix inequalities (8.19) and (8.20) are equivalent to LMIs (8.14) and
(8.15), respectively. In regard to the stableness of the closed-loop system, we
notice that both (8.19) and (8.20) imply that (A+ BF )T P + P (A+ BK) < 0,
which further implies that the closed-loop system is stable. This completes
the proof. 
Notice that the argument in the end of the above proof also applies to
Theorems 8.3.1, 8.4.1, and 8.5.1 given later on. Hence we will not repeat it
in the proof of these theorems.
In the decoupling controller design for practical plants, the magnitude of
parameters βi , i = 1, 2, . . . , m, can be readily estimated from Remark 8.2.1,
but it is hard to infer the magnitude of parameter γ. Therefore we formulate
the near-decoupling controller design (NDCD) in the case of state-feedback
(SF) as the following optimization problem.
276 8. Near-Decoupling

Problem NDCD-SF:
minimize γ
subject to LMIs Q > 0, (8.14), (8.15).
The above problem is a standard eigenvalue problem (EVP) in LMI lan-
guage (Boyd et al., 1994), hence its numerical solutions can be found through
LMI toolbox of Matlab very efficiently (Gahinet et al., 1995).

8.3 Robust Near-Decoupling: State Feedback

Now consider the case that there are some parameter perturbations in the
system matrix A. We assume that the perturbations are structured as follows
A = A0 + Al Ψ Ar , (8.21)
n1 ×n2
Ψ ∈ Ψ̆ := {Ψ ∈ R : kΨ k2 ≤ 1},
where A0 , a constant matrix, is the nominal parameter of A, Ψ denotes the
uncertainty, Al and Ar , constant matrices with appropriate dimensions, are
used to describe how the uncertainty Ψ enters the system matrix A. Notice
that any uncertainties with bounded norm can be described by the form
(8.21).
Theorem 8.3.1. Consider uncertain system (8.2) with matrix A described
by (8.21). If there are matrices Q > 0, F and numbers εij > 0 (i, j =
1, 2, . . . , p) such that the following LMIs
 
−QAT0 − A0 Q − F T B T − BF − εii Al ATl −Bi + (QCiT + F T DiT )Dii QATr
 T T
−Bi + Dii (Ci Q + Di F ) 2 T
−βi I + Dii Dii 0  > 0,
Ar Q 0 εii I
∀i ∈ {1, 2, . . . , m}, (8.22)

 
QAT0 + A0 Q + F T B T + BF + εij Al ATl Bj QCiT + F T DiT QATr
T T
 Bj −γI Dij 0  < 0,
 
Ci Q + Di F Dij −γI 0
Ar Q 0 0 −εij I
∀i, j ∈ {1, 2, . . . , m}, i 6= j, (8.23)

hold, then system (8.2) is nearly state-feedback decouplable. The state-

feedback law and closed-loop performance indices are given by (8.16) and
(8.17)-(8.18) respectively.

To prove this theorem, we need the following lemma.

Lemma 8.3.1. Let Φl , Φr , Υ be real matrices of appropriate dimensions with

Υ satisfying kΥ k2 ≤ 1. Then for any real number ε > 0 we have

Φl Υ Φr + ΦTr Υ T ΦTl ≤ εΦl ΦTl + ε−1 ΦTr Φr . (8.24)

 8.3 Robust Near-Decoupling: State Feedback 277

Proof. The assumption kΥ k2 ≤ 1 leads to I − Υ Υ T ≥ 0, which again yields

the following inequality
 
I Υ
≥0
ΥT I

by applying Schur complements (Boyd et al., 1994). Then for any two real
numbers ε1 and ε2 we will have
  
  I Υ ε1 ΦTl
ε1 Φl −ε2 ΦTr ≥ 0. (8.25)
ΥT I −ε2 Φr

If ε1 ε2 = 1, the inequality (8.25) will be of the following form

ε21 Φl ΦTl + ε22 ΦTr Φr − Φl Υ Φr − ΦTr Υ T ΦTl ≥ 0,

which is equivalent to inequality (8.24). 

Proof of theorem 8.3.1. It follows from Theorem 8.2.1 that if there are ma-
trices Q > 0 and F such that the following inequalities
 
−QAT − AQ − F T B T − BF −Bi + (QCiT + F T DiT )Dii
> 0,
−BiT + Dii
T
(Ci Q + Di F ) −βi2 I + Dii
T
Dii
∀i ∈ {1, 2, . . . , m}, ∀Ψ ∈ Ψ̆ ; (8.26)
 
QAT + AQ + F T B T + BF Bj QCiT + F T DiT
 BjT −γI T
Dij  < 0,
Ci Q + Di F Dij −γI
∀i, j ∈ {1, 2, . . . , m}, i 6= j, ∀Ψ ∈ Ψ̆ (8.27)
hold, then system (8.2) is nearly state-feedback decouplable with performance
requirement (8.17)-(8.18) being satisfied. The difficulty lies in the fact that
the above inequalities should be satisfied for all possible uncertainty Ψ. Now
we try to remove the term Ψ in the above inequalities. In view of Lemma
8.3.1, we have
QAT + AQ = QAT0 + A0 Q + QATr Ψ T ATl + Al Ψ Ar Q
≤ QAT0 + A0 Q + ε−1 QATr Ar Q + εAl ATl , ∀ε ∈ (0, +∞).
Therefore inequalities (8.26) and (8.27) hold if the following inequalities
 
−(QAT0 + A0 Q + ε−1 T T T T T T T
ii QAr Ar Q + εii Al Al ) − F B − BF −Bi + (QCi + F Di )Dii > 0,
−BiT + Dii
T
(Ci Q + Di F ) −βi2 I + Dii
T
Dii
∀i ∈ {1, 2, . . . , m}; (8.28)
 
QAT0 + A0 Q + ε−1 T T T T T T T
ij QAr Ar Q + εij Al Al + F B + BF Bj QCi + F Di
 BjT
−γI T
Dij  < 0,
Ci Q + Di F Dij −γI
∀i, j ∈ {1, 2, . . . , m}, i 6= j (8.29)
278 8. Near-Decoupling

are satisfied for some positive numbers εij (i, j = 1, 2, . . . , m). It follows from
Schur complements that inequalities (8.28) and (8.29) are equivalent to LMIs
(8.22) and (8.23), respectively. Thus the proof is completed. 
Similarly, we can formulate the robust near-decoupling controller design
(RNDCD) in the case of state-feedback as the following optimization problem.
Problem RNDCD-SF:
minimize γ
subject to LMIs Q > 0, (8.22), (8.23).
In both problems NDCD-SF and RNDCD-SF, the minimization proce-
dure will give a very large Q, making Q−1 ill-conditioned. Therefore for these
two problems, we impose the following additional constraint

Q < µI

on the corresponding problems in the practical controller computation, where

µ is a given positive number.

8.4 Near-Decoupling: Dynamic Output Feedback

In many cases, all state variables are not available. Thus it is necessary to
develop a counterpart of the results of the preceding two sections based on
output feedback. In the two sections to be followed, we will try to deal with
this issue. Our controller is described by the following dynamical output
feedback law:

ż = AK z + BK (y − Du),
(8.30)
u = CK z + DK (y − Du) + v,

i.e.,

ż = AK z + BK Cx,
(8.31)
u = CK z + DK Cx + v.

We suppose that

|DK D + I| 6= 0.

Our basic idea is to use the results already obtained hereto to develop a
method for the design of controller (8.30). To proceed, we write the closed-
loop system consisting of (8.2) and (8.31) as follows
      

 ẋ A + BDK C BCK x B
 ż = BK C AK z
+
0
v,
  (8.32)

   x
 y = C + DDK C DCK + Dv.
z
 8.4 Near-Decoupling: Dynamic Output Feedback 279

Let
   
A + BDK C BCK B  
Ā = , B̄ = , C̄ = C + DDK C DCK , D̄ = D.
BK C AK 0

Then it follows from Lemma 8.2.1 and Corollary 8.2.1 that system (8.32) is
near-decoupled with performance requirement (8.17)-(8.18) being satisfied if
there exists a matrix P̄ > 0 such that the following matrix inequalities
 
−ĀT P̄ − P̄ Ā −P̄ B̄i + C̄iT D̄ii
> 0, ∀i ∈ {1, 2, . . . , m}, (8.33)
T
−B̄iT P̄ + D̄ii T
C̄i −βi2 I + D̄ii D̄ii

 
ĀT P̄ + P̄ Ā P̄ B̄j C̄iT
 B̄jT P̄ −γI D̄ij T 
< 0, ∀i, j ∈ {1, 2, . . . , m}, i 6= j (8.34)
C̄i D̄ij −γI

hold.
Due to the fact that the matrix Ā also includes the design parameters
AK , BK , CK and DK , (8.33) and (8.34) are bilinear matrix inequalities in the
unknowns P̄ , AK , BK , CK and DK , which is very difficulty to solve. Thanks
to the pioneer work by Chilali, Gahinet and Scherer (Chilali and Gahinet,
1996; Scherer et al., 1997), both (8.33) and (8.34) can be transformed into
two LMIs by taking a change of variables. In the following, we will use the
method developed in (Chilali and Gahinet, 1996; Scherer et al., 1997) to solve
our problem.
Let us partition P̄ and P̄ −1 as
   
Y N −1 X M
P̄ = , P̄ = , (8.35)
NT ? MT ?

where X and Y are n × n positive definite matrices, M and N are n × k

matrices with k being the order of the controller (8.30), and ? denotes the
matrices in which we are not interested. Clearly M and N should satisfy the
following relationship

M N T = I − XY. (8.36)

It will be clear later that X and Y , together with other matrices, are design
variables which make the required performance indices be satisfied, while M
and N are free parameters, the unique requirement for which is to satisfy
equation (8.36). Define
   
X I I Y
Π1 = , Π2 = .
MT 0 0 NT

It is clear that

P̄ Π1 = Π2 .
280 8. Near-Decoupling

Now define the change of controller variables as follows:

 T T
 Â := Y (A + BDK C)X + N BK CX + Y BCK M + N AK M


B̂ := Y BDK + N BK
(8.37)

 Ĉ := DK CX + CK M T

D̂ := DK
Then a short calculation yields the following identities
  

 T T AX + B Ĉ A + B D̂C

 Π1 P̄ ĀΠ1 = Π2 ĀΠ1 = ,


   Â Y A + B̂C

 B
Π1T P̄ B̄ = Π2T B̄ = ,
 Y B  (8.38)



 C̄Π1 = CX + DĈ C +DD̂C ,



 Π1T P̄ Π1 = Π1T Π2 = X I .

I Y
The partition of P̄ and P̄ −1 defined in (8.35), the new controller vari-
ables identified in (8.37) and the relationship shown in (8.38) are developed
in (Chilali and Gahinet, 1996; Scherer et al., 1997), which make it easy to
transform many problems of dynamic output feedback based controller de-
sign to LMI related problems. In the sequel, we make another try along this
direction while solving our problems.
Multiplying the left- and right-hand-sides of inequality (8.33) by block-
diag {Π1T , I} and its transpose respectively yields
   
Π1T 0 −ĀT P̄ − P̄ Ā −P̄ B̄i + C̄iT D̄ii Π1 0
0 I −B̄iT P̄ + D̄ii
T T
C̄i −βi2 I + D̄ii D̄ii 0 I
 
−AX − XAT − B Ĉ − Ĉ T B T −A − B D̂C − ÂT −Bi + (Ci X + Di Ĉ)T Dii
= T
−(A + B D̂C) − Â −Y A − A Y − B̂C − C B̂ −Y Bi + (Ci + Di D̂C) Dii 
T T T T
T T
−Bi + Dii (Ci X + Di Ĉ) T
−BiT Y + Dii (Ci + Di D̂C) T
−βi2 I + Dii Dii
> 0, ∀i ∈ {1, 2, . . . , m}; (8.39)
Multiplying the left- and right-hand-sides of inequality (8.34) by block-
diag {Π1T , I, I} and its transpose respectively yields
  T  
Π1T 0 0 Ā P̄ + P̄ Ā P̄ B̄j C̄iT Π1 0 0
 0 I 0   B̄jT P̄ −γI D̄ij T 
0 I 0
0 0I C̄i D̄ij −γI 0 0I
 
AX + XAT + B Ĉ + Ĉ T B T A + B D̂C + ÂT Bj (Ci X + Di Ĉ)T
 T
(A + B D̂C) + Â Y A + AT Y + B̂C + C T B̂ T Y Bj (Ci + Di D̂C)T 
=



BjT BjT Y −γI T
Dij
Ci X + Di Ĉ Ci + Di D̂C Dij −γI
< 0, ∀i, j ∈ {1, 2, . . . , m}, i 6= j. (8.40)
In the above derivation, we have used the following convention:
 
Bi  
B̄i = , C̄i = Ci + Di DK C Di CK ,
0
 
D̄ij = Dij , D̄i = D̄i1 D̄i2 · · · D̄ip .
 8.4 Near-Decoupling: Dynamic Output Feedback 281

Thus, similar to equation (8.38), one has the following identities

 
T T Bi
Π1 P̄ B̄i = Π2 B̄i = ,
Y Bi
 
C̄i Π1 = Ci X + Di Ĉ Ci + Di D̂C .
Notice that inequalities (8.39) and (8.40) are linear in the unknown vari-
ables X, Y, Â, B̂, Ĉ and D̂. Once the solutions for these variables are
found, the original controller parameters can be recovered from (Chilali and
Gahinet, 1996; Scherer et al., 1997)


 DK = D̂

CK = (Ĉ − DK CX)M −T

 B = N −1 (B̂ − Y BDK )
 K
AK = N −1 [Â − Y (A + BDK C)X − N BK CX − Y BCK M T ]M −T
(8.41)
with the assumption that both M and N are square. The nonsingularness of
M and N comes from the fact that X − Y −1 > 0, which can be inferred from
the last equation of equation (8.38).
We summarize the above results into the following theorem.
Theorem 8.4.1. System (8.2) is near-decoupled under dynamic output feed-
back controller (8.30) with performance (8.17)-(8.18) if there exist matrices
X, Y, Â, B̂, Ĉ and D̂ such that the following LMIs
 
−AX − XAT − B Ĉ − Ĉ T B T −A − B D̂C − ÂT −Bi + (Ci X + Di Ĉ)T Dii
 T
−(A + B D̂C) − Â −Y A − A Y − B̂C − C B̂ −Y Bi + (Ci + Di D̂C) Dii  > 0,
T T T T
T
−BiT + Dii (Ci X + Di Ĉ) T
−BiT Y + Dii (Ci + Di D̂C) T
−βi2 I + Dii Dii
∀i ∈ {1, 2, . . . , m}, (8.42)
 
AX + XAT + B Ĉ + Ĉ T B T A + B D̂C + ÂT Bj (Ci X + Di Ĉ)T
 T
(A + B D̂C) + Â Y A + AT Y + B̂C + C T B̂ T T 
Y Bj (Ci + Di D̂C) 

 BjT BjT Y −γI T
Dij  < 0,
Ci X + Di Ĉ Ci + Di D̂C Dij −γI
∀i, j ∈ {1, 2, . . . , m}, i 6= j, (8.43)
 
X I
>0 (8.44)
I Y

hold. In this case, the controller parameters are given by (8.41).

As before, the near-decoupling controller design in the case of output-
feedback (OF) can be formulated as the following optimization problem.
Problem NDCD-OF:
minimize γ
subject to LMIs (8.42), (8.43), (8.44).
282 8. Near-Decoupling

8.5 Robust Near-Decoupling: Dynamic Output

Feedback
Now suppose that the system matrix A is uncertain and of the structure
defined by equation (8.21). Again we try to find a dynamic output feedback
controller of the form (8.30) to achieve the objective of near-decoupling. To
this end, we have
Theorem 8.5.1. Consider uncertain system (8.2) with matrix A described
by (8.21). If there exist matrices X, Y, Â, B̂, Ĉ, D̂ and positive numbers ε3
and ε4 such that the following matrix inequalities
 
−A0 X − XAT0 − B Ĉ − Ĉ T B T −A0 − B D̂C − ÂT0 −Bi + (Ci X + Di Ĉ)T Dii Al XATr
 T
−(A0 + B D̂C) − Â0 T T T T
−Y A0 − A0 Y − B̂C − C B̂ −Y Bi + (Ci + Di D̂C) Dii Y Al Ar  T 

 0 
−BiT + Dii  > 0,
T
 (Ci X + Di Ĉ) −BiT Y + Dii
T
(Ci + Di D̂C) −βi2 I + Dii
T
Dii 0
 AlT
ATl Y 0 ε3 I 0 
−1
Ar X Ar 0 0 ε3 I

 ∀i ∈ {1, 2, . . . , m}, (8.45)


A0 X + XAT0 + B Ĉ + Ĉ T B T A0 + B D̂C + ÂT0 Bj (Ci X + Di Ĉ)T Al XATr
 (A + B D̂C) T
+ Â Y A + AT0 Y + B̂C + C T B̂ T Y Bj (Ci + Di D̂C)T Y Al ATr 
 0 0 0 
 B T
BjT Y −γI T
Dij 0 0 
 j <0
 Ci X + Di Ĉ Ci + Di D̂C Dij −γI 0 0 
 
 ATl ATl Y 0 0 −ε4 I 0 
Ar X Ar 0 0 0 −ε−1 4 I

  ∀i, j ∈ {1, 2, . . . , m}, i 6= j, (8.46)

X I
>0 (8.47)
I Y
hold, then system (8.2) is nearly output-feedback decouplable. The output-
feedback law and closed-loop performance indices are given by (8.30)-(8.41)
and (8.17)-(8.18) respectively.
Proof. Define
Â0 := Y (A0 + BDK C)X + N BK CX + Y BCK M T + N AK M T ,
 := Y (A + BDK C)X + N BK CX + Y BCK M T + N AK M T
= Â0 + Y Al Ψ Ar X,
B̂, Ĉ, D̂ := as in equation (8.37).
Following the same lines of the argument as in Theorem 8.4.1, we can conclude
that if the following inequalities
 
−AX − XAT − B Ĉ − Ĉ T B T −A − B D̂C − ÂT −Bi + (Ci X + Di Ĉ)T Dii
 −(A + B D̂C)T − Â −Y A − AT Y − B̂C − C T B̂ T −Y Bi + (Ci + Di D̂C)T Dii 
−BiT + Dii T
(Ci X + Di Ĉ) −BiT Y + DiiT
(Ci + Di D̂C) −βi2 I + Dii
T
Dii
 T T T T

−A0 X − XA0 − B Ĉ − Ĉ B −A0 − B D̂C − Â0 −Bi + (Ci X + Di Ĉ)T Dii
=  T
−(A0 + B D̂C) − Â0 −Y A0 − A0 Y − B̂C − C B̂ −Y Bi + (Ci + Di D̂C) Dii 
T T T T

−BiT + Dii T
(Ci X + Di Ĉ) −BiT Y + DiiT
(Ci + Di D̂C) −βi2 I + Dii
T
Dii
 T T

Al Ψ Ar X + X(Al Ψ Ar ) Al Ψ Ar + (Y Al Ψ Ar X) 0
−  (Al Ψ Ar )T + Y Al Ψ Ar X Y Al Ψ Ar + (Al Ψ Ar )T Y 0 
0 0 0
> 0, ∀i ∈ {1, 2, . . . , m}, ∀Ψ ∈ Ψ̆ , (8.48)
 8.5 Robust Near-Decoupling: Dynamic Output Feedback 283

 
AX + XAT + B Ĉ + Ĉ T B T A + B D̂C + ÂT Bj (Ci X + Di Ĉ)T
 (A + B D̂C)T + Â Y A + AT Y + B̂C + C T B̂ T T 
Y Bj (Ci + Di D̂C) 

 BjT BjT Y −γI T
Dij 
Ci X + Di Ĉ Ci + Di D̂C Dij −γI
 
A0 X + XAT0 + B Ĉ + Ĉ T B T A0 + B D̂C + ÂT0 Bj (Ci X + Di Ĉ)T
 (A0 + B D̂C)T + Â0 Y A0 + AT0 Y + B̂C + C T B̂ T T 
Y Bj (Ci + Di D̂C) 
= 
 BjT BjT Y −γI T
Dij 
Ci X + Di Ĉ Ci + Di D̂C Dij −γI
 
Al Ψ Ar X + X(Al Ψ Ar )T Al Ψ Ar + (Y Al Ψ Ar X)T 0 0
 (Al Ψ Ar )T + Y Al Ψ Ar X Y Al Ψ Ar + (Al Ψ Ar )T Y 0 0 
+


0 0 0 0
0 0 00
< 0, ∀i, j ∈ {1, 2, . . . , m}, i 6= j, ∀Ψ ∈ Ψ̆ (8.49)
together with (8.47) hold, system (8.2) is near-decoupled and the closed-loop
performance indices are given by (8.17)-(8.18). In view of Lemma 8.3.1, we
have
 
Al Ψ Ar X + X(Al Ψ Ar )T Al Ψ Ar + (Y Al Ψ Ar X)T 0
 (Al Ψ Ar )T + Y Al Ψ Ar X Y Al Ψ Ar + (Al Ψ Ar )T Y 0 
0 0 0
   T
Al Ψ Ar X Al Ψ Ar 0 Al Ψ Ar X Al Ψ Ar 0
=  Y Al Ψ Ar X Y Al Ψ Ar 0  +  Y Al Ψ Ar X Y Al Ψ Ar 0 
0 0 0 0 0 0
    T
Al   Al  
=  Y Al  Ψ Ar X Ar 0 +  Y Al  Ψ Ar X Ar 0 
0 0
   
Al   XATr  
−1 
≤ ε3 Y Al  ATl ATl Y 0 + ε3  ATr  Ar X Ar 0 , ∀ε3 ∈ (0, +∞)
0 0
and
 
Al Ψ Ar X + X(Al Ψ Ar )T Al Ψ Ar + (Y Al Ψ Ar X)T 0 0
T T
 (Al Ψ Ar ) + Y Al Ψ Ar X Y Al Ψ Ar + (Al Ψ Ar ) Y 0 0 
 
0 0 00
0 0 00
   
Al XATr
 Y A l
   Ar  
T 
≤ ε−1 T T
4  0  Al Al Y 0 0 + ε4   Ar X Ar 0 0 , ∀ε4 ∈ (0, +∞).
0
0 0
Therefore, if the following inequalities
 
−A0 X − XAT0 − B Ĉ − Ĉ T B T −A0 − B D̂C − ÂT0 −Bi + (Ci X + Di Ĉ)T Dii
 T
−(A0 + B D̂C) − Â0 −Y A0 − A0 Y − B̂C − C B̂ −Y Bi + (Ci + Di D̂C) Dii 
T T T T

−BiT + DiiT
(Ci X + Di Ĉ) −BiT Y + Dii
T
(Ci + Di D̂C) −βi2 I + Dii
T
Dii
   T

Al   XA r  
−ε−1
3
 Y Al  ATl ATl Y 0 − ε3  ATr  Ar X Ar 0
0 0
> 0, ∀i ∈ {1, 2, . . . , m}, (8.50)
284 8. Near-Decoupling

 
A0 X + XAT0 + B Ĉ + Ĉ T B T A0 + B D̂C + ÂT0 Bj (Ci X + Di Ĉ)T
 T
(A0 + B D̂C) + Â0 Y A0 + AT0 Y + B̂C + C T B̂ T Y Bj (Ci + Di D̂C)T 
 
 BjT BjT Y −γI T
Dij 
Ci X + Di Ĉ Ci + Di D̂C Dij −γI
   
Al XATr
 Y Al   T T   ATr   
+ε−1    
4  0  Al Al Y 0 0 + ε4  0  Ar X Ar 0 0

0 0
< 0, ∀i, j ∈ {1, 2, . . . , m}, i 6= j (8.51)

hold, so do inequalities (8.48) and (8.49). By Schur complements, inequalities

(8.50) and (8.51) is equivalent to inequalities (8.45) and (8.46), respectively.
This completes the proof. 
Different from the problems dealt with in the preceding sections, inequal-
ities (8.45) and (8.46) are not linear in the unknowns ε3 and ε4 , respectively.
Thus we cannot use standard LMI toolbox of Matlab to solve these two in-
equalities. An important characteristic of these two inequalities is that they
reduce to LMIs once upon ε3 and ε4 are fixed. Therefore one possible way is
to solve inequalities (8.45), (8.46) and (8.47) by sweeping all possible ε3 and
ε4 , but the difficulty is where one should start his sweeping. From inequalities
(8.50) and (8.51), we can see that ε3 and ε4 in inequalities (8.45) and (8.46)
are of the similar rule and structure as the “communicating variables” in the
problems studied in (Zheng et al., 2002; Zheng et al., to appear). In (Zheng
et al., 2002; Zheng et al., to appear), an optimal estimate for the communi-
cating variables is given, which facilitates us to solve the relevant problems.
In the sequel, we will give an estimate for the initial values of ε3 and ε4 based
on the method in (Zheng et al., 2002; Zheng et al., to appear).
From inequalities (8.50) and (8.51) and following the same argument as
that in (Zheng et al., 2002; Zheng et al., to appear), one can see that the
most likely values for ε3 and ε4 which make inequalities (8.45) and (8.46)
hold are given by
v   
u Al 
u 
u  Y A  AT AT Y 0 
u tr l l l s
u
u 0 tr(Al ATl ) + tr(Y Al ATl Y )
ε∗3 = u    = .
u XATr  tr(ATr Ar ) + tr(XATr Ar X)
u 
t tr  ATr  Ar X Ar 0 
0
:= (X, Y ) (8.52)
ε∗4 = (X, Y ).
Since both ε∗3 and ε∗4 are functions of the unknowns X and Y , it is still
difficult to solve our problem. Because matrix inequalities (8.45), (8.46) and
(8.47) should have solutions when there are no parameter perturbations, we
can reasonably use the matrices X and Y that solve problem NDCD-OF to
 8.6 A Numerical Example 285

calculate ε∗3 and ε∗4 , which are then used as the initial values for ε3 and ε4
respectively in the sweeping procedure.
Based upon the above discussion, we can adopt the following procedure to
solve robust near-decoupling controller design (RNDCD) problem for output
feedback case.

Design procedure for RNDCD-OF problem:

Step 0. Initial data: System’s state space realization (A0 , B, C, D, Al , Ar ),
performance indices βi (i = 1, 2, . . . , p).
Step 1. Solve problem NDCD-OF. Denote by X0 and Y0 its solution for the
variables X and Y , respectively.
Step 2. Calculate ε0 = (X0 , Y0 ) according to equation (8.52).
Step 3. for ε = ε0 down to εf 1 or up to εf 2 , where εf 1 and εf 2 are two
positive numbers, solve the following optimization problem:
minimize γ
subject to LMIs (8.45), (8.46), (8.47), in which ε3 and ε4
are set to be ε.
Notice that in the above procedure we set ε3 = ε4 to avoid a two dimen-
sional sweeping, which would introduce additional conservativeness for the
original problem.
As pointed out in (Chilali and Gahinet, 1996; Scherer et al., 1997), a min-
imization procedure often make the matrix I −XY nearly singular. This phe-
nomenon is also evidenced while we numerically solve problems NDCD-OF
and RNDCD-OF. Therefore, similar to (Chilali and Gahinet, 1996; Scherer
et al., 1997), we use the following matrix inequality
 
X ηI
> 0, (8.53)
ηI Y

to substitute matrix inequalities (8.44) and (8.47) in the corresponding prob-

lems, where η is a given number satisfying η  1.

8.6 A Numerical Example

Now let us consider the following 2 × 2 system

ẋ = Ax + Bu, y = Cx + Du

with
   
−1 + ∆ 0 0 1 0
A= 0 −2 + 2∆ 0 , B =  2 3,
0 0 −3 + 3∆ −3 −3
   
100 10
C= , D= ,
111 01
286 8. Near-Decoupling

where ∆ is an uncertain parameter varying in the closed interval [−0.1, 0.1].

The nominal value for matrix A is
 
−1 0 0
A0 =  0 −2 0 
0 0 −3
and the uncertainty can be expressed in the form of (8.21) with
   
100 0.1 0 0
Al =  0 2 0  , Ar =  0 0.1 0  .
003 0 0 0.1
Its nominal transfer matrix is
" s+2 #
s+1 0
H0 (s) = 4s+6 s2 +5s+9
.
(s+1)(s+2)(s+3) (s+2)(s+3)

The Bode diagrams for this system (nominal part) are illustrated in Fig. 8.1.
Using the methods developed in the preceding sections, we can obtain the
near-decoupling controllers for SF and OF cases and achievable performance,
as summarized in Table 8.1.

Table 8.1. Near-decoupling Controllers and Their Performance

Problem Design param. Perfor. guaranteed Feedback matrices
β1 = β2 = 0.95  
−0.9998 0.0018 0.0023
NDCD-SF γ = 3.61 × 10−4 K =
µ = 104 −0.9993 −1.0001 −0.9997

β1 = β2 = 0.95,  
−1.0009 0.0006 0.0009
RNDCD-SF γ = 4.43 × 10−4 K =
µ = 104 −0.9993 −1.0001 −0.9998

 
0.0103 6.6284 −6.6386
AK =  0.0159 6.6257 −6.6416  × 108
0.0159 6.6257 −6.6416
 
−1.0485 −2.1362
BK =  −0.5438 −1.8575  × 108
NDCD-OF β1 = β2 = 0.95 γ = 6.7166 × 10−4 −0.5438 −0.5748
η = 104  
0.0223 −0.0105 −0.0118
CK =
7.2e{−9} 2.0e{−5} −2.0e{−5}

DK = D
 
−2.2333 −1.2025 −0.5697
AK =  −0.0930 −2.1850 −2.7781  × 106
−0.0930 −2.1850 −2.7781
 
−5.9078 −2.4216
β1 = β2 = 0.95 γ = 0.001 at BK =  1.7431 −4.1658  × 108
NDCD-ROF 1.7431 −4.1658
η = 10 ε3 = ε4 = 0.001
 
0.0063 −0.0002 −0.0030
CK =
0.0031 0.0114 0.0111

DK = D

The Bode diagrams for the closed-loop system (we choose ∆ = 0.1) are
illustrated in Figs. 8.2, 8.3, 8.4 and 8.5 for the cases NDCD-SF, RNDCD-SF,
NDCD-OF and RNDCD-OF respectively.
 8.6 A Numerical Example 287

Comparing Table 8.1 and Figs. 8.2, 8.3, 8.4 and 8.5, we can see that the
performance indices guaranteed by our methods are somewhat conservative,
especially for the case where there are parameter perturbations. This is not
strange since the performance index γ shown in Table 8.1 is guaranteed for
all possible uncertain models.

2.2 1

0.8
2

0.6

1.8
0.4

0.2
1.6
|H11|

|H12|

1.4
−0.2

−0.4
1.2

−0.6

1
−0.8

0.8 −1
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(a) H11 (s) (b) H12 (s)

1.2 1.6

1.5
1

1.4
0.8

1.3
|H21|

|H22|

0.6

1.2

0.4
1.1

0.2
1

0 0.9
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(c) H21 (s) (d) H22 (s)

Fig. 8.1. Bode diagrams of the open-loop transfer matrices.

288 8. Near-Decoupling

−4
1.0001 x 10
4

1 3.5

0.9999 3

2.5
0.9998
|
cl,11

|Hcl,12|
2
|H

0.9997

1.5

0.9996
1

0.9995
0.5

0.9994 0
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(a) Hcl,11 (s) (b) Hcl,12 (s)

−4 −3
x 10 x 10
3.5 10.05

10
3

9.95
2.5

9.9
|Hcl,22|−0.99

2
|Hcl,21|

9.85

1.5
9.8

1
9.75

0.5
9.7

0 9.65
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(c) Hcl,21 (s) (d) Hcl,22 (s)

Fig. 8.2. Bode diagrams of the closed-loop transfer matrices: ADCD-SF.

8.7 Concluding Remarks

In this chapter, we have proposed the new concept of near-decoupling con-

trol. Fundamental lemmas for solving near-decoupling control problem are
established in the form of linear matrix inequalities. Design methods on the
near-decoupling controllers have been presented for four general cases: (i)
exact models and state feedback, (ii) uncertain models and state feedback,
(iii) exact models and dynamic output feedback, and (iv) uncertain models
and dynamic output feedback. In the case of uncertain models, we assume
that the parameter uncertainties are norm-bounded and may be of some
structure properties. All these results are described in LMI language, which
 8.7 Concluding Remarks 289

−4
1.0001 x 10
2

1 1.8

1.6
0.9999

1.4
0.9998
1.2
|
cl,11

|Hcl,12|
0.9997
1
|H

0.9996 0.8

0.6
0.9995

0.4

0.9994
0.2

0.9993 0
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(a) Hcl,11 (s) (b) Hcl,12 (s)

−4 −3
x 10 x 10
4 10.05

3.5 10

3 9.95

2.5 9.9
|Hcl,22|−0.99
|Hcl,21|

2 9.85

1.5 9.8

1 9.75

0.5 9.7

0 9.65
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(c) Hcl,21 (s) (d) Hcl,22 (s)

Fig. 8.3. Bode diagrams of the closed-loop transfer matrices: RADCD-SF.

makes the numerical solutions of near-decoupling controller design problem

easily tractable. The results given in this chapter provides a viable way for
solving the problem of robust decoupling control, which is of conspicuous
significance, but unfortunately, has long remained rarely studied in control
theory.
While we have witnessed the effectiveness of the approach developed here
through the illustrative example, we should admit that there exist two main
limitations in our approach, which are
(i) The matrix D cannot become vanished. This is due to the specification
defined in equation (8.3). If we change this equation as follows
290 8. Near-Decoupling

−6 −7
x 10 x 10
3.9 1.4

3.85
1.2

3.8
1

3.75
0.8
|Hcl,11|−1.0

|Hcl,12|
3.7

0.6
3.65

0.4
3.6

0.2
3.55

3.5 0
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(a) Hcl,11 (s) (b) Hcl,12 (s)

−7 −5
x 10 x 10
1.4 10.004

10.002
1.2

10
1

9.998
|−0.9999

0.8
|Hcl,21|

9.996
cl,22

0.6
|H

9.994

0.4
9.992

0.2
9.99

0 9.988
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(c) Hcl,21 (s) (d) Hcl,22 (s)

Fig. 8.4. Bode diagrams of the closed-loop transfer matrices: ADCD-OF.

σ(Hii (jω)) ≥ βi , ∀i ∈ {1, 2, . . . , p}, ∀ω ∈ B, (8.54)

where B is a limited band of frequencies in which the designer is inter-

ested, the matrix D can then be allowed to be zero. The difficulty lies in
the shortage of effective methods which convert specification (8.54) to a
numerical tractable formula concerning system matrices.
(ii) The decoupling controller is limited to the class of (8.5), which decreases
the freedom of controller design parameters. In practice, it is desirable
that the controller is of the following general form

u = Kx + Gv.
 8.7 Concluding Remarks 291

−7 −8
x 10 x 10
10.4 1.4

1.2
10.2

1
10

0.8
|Hcl,11|−1.0

|Hcl,12|
9.8

0.6

9.6
0.4

9.4
0.2

9.2 0
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(a) Hcl,11 (s) (b) Hcl,12 (s)

−7 −7
x 10 x 10
5.5 4

5 3.5

4.5 3

4 2.5
|Hcl,22|−1.0
|Hcl,21|

3.5 2

3 1.5

2.5 1

2 0.5

1.5 0
−2 −1 0 1 2 3 −2 −1 0 1 2 3
10 10 10 10 10 10 10 10 10 10 10 10
Frequency (rad/sec) Frequency (rad/sec)

(c) Hcl,21 (s) (d) Hcl,22 (s)

Fig. 8.5. Bode diagrams of the closed-loop transfer matrices: RADCD-OF.

It is observed that the main rule of matrix G is to statically change the

transfer matrices of the closed-loop systems. We hope that this limitation
can be partly overcome in the pre-design stage by considering the input-
output pairs to be decoupled.
Finally we would like to point out that the concept of near-decoupling
design proposed in (Cheng et al., 1995) is of some similarity with the concept
of robust near-decoupling proposed in this chapter, but the design procedures
are completely different between this chapter and the references (Cheng et al.,
1995), where the “decoupling” controller is designed still based on the nominal
292 8. Near-Decoupling

model of the plants, but their objective is to nearly, instead of completely,

decouple the system.

8.8 Notes and References

A literature survey will reveal that nominal decoupling problem has now
been well studied. In sharp contrast to other control topics where robustness
has been extensively investigated, robust decoupling problem for uncertain
systems has rarely been reported in the literature. This phenomena is un-
satisfactory since parameter perturbations or deviations from their nominal
values are inevitable in real plants, especially in industrial process control.
Intuitively, exact decoupling is very difficult to achieve owing to the presence
of plant parameter perturbations. In practice, it is often the case that one first
designs a decoupling controller (decoupler) for the nominal model of the real
plant, put the decoupler as the inner loop controller and then on-line tune
the outer loop controller to make the overall system performance be roughly
satisfied. Reference (Wu et al., 1998) describes such a procedure. In robot
control systems, the calculated torque method combined with gain schedul-
ing (Ito and Shiraishi, 1997) also belongs to this approach. It is believed that
near or robust decoupling problem deserves more attention from researchers.
9. Dynamic Disturbance Decoupling

The attenuation of load disturbance is always of a primary concern for any

control system design, and is even the ultimate objective for process con-
trol, where the set-point may be kept unchanged for years. As such, this and
next chapters will deal with decoupling disturbances from the plant output in
dynamic and asymptotic senses, respectively. The dynamic disturbance de-
coupling problem is to find a control scheme such that in the resulting control
system the transfer matrix function from the disturbance to the controlled
output is zero for all frequencies, i.e. there is no effect of the disturbance
on the controlled output. If the disturbance is measurable, the feedforward
compensation scheme can be employed to eliminate its effect on the system
output, which will be the topic of Section 1 of this chapter. For unmeasur-
able disturbances which are more often encountered in industry, feedback
control has to be adopted. However, there will be inevitably a design trade-
off between the set-point response and disturbance rejection performance. To
alleviate this problem, a control scheme, called the disturbance observer, is
introduced and it acts as an add-on mechanisms to the conventional feed-
back system. The disturbance observer estimates the equivalent disturbance
as the difference between the actual process output and the output of the
nominal model. The estimate is then fed to a process model inverse to pro-
duce an extra control effort which can compensate for the disturbance effect
on the output. The disturbance observer will be addressed in Section 2 of this
chapter. Stability is a crucial issue and will be considered in both schemes.

9.1 Measurable Disturbance: Feedforward Control

In this section, supposing measurable disturbances, we aim to solve distur-
bance decoupling problem with stability via disturbance feedforward control
in connection with the normal output feedback configuration. Let a given
system be described by a transfer matrix model
y(s) = Gu (s)u(s) + Gd (s)d(s), (9.1)
ym (s) = Gm (s)y(s), (9.2)
where y(s) is the m-dimensional controlled output, u(s) the l-dimensional
control input, d(s) the p-dimensional disturbance, ym (s) q-dimensional mea-
surement, and the proper rational matrices Gu , Gm and Gd with appropriate
Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 293-325, 2003.
Springer-Verlag Berlin Heidelberg 2003
294 9. Dynamic Disturbance Decoupling

dimensions represent the plant model, sensor model and disturbance channel
model, respectively. Suppose that both ym and d are accessible. A general lin-

Fig. 9.1. Control Scheme for Decoupling Measurable Disturbances

ear compensator which makes use of all the available information is described
by
u(s) = −Gf (s)d(s) − Gb (s)ym (s), (9.3)
where Gf and Gb are proper rational matrices to be designed for distur-
bance decoupling. This compensator can be thought as a combination of a
disturbance feedforward part and a output feedback part. The overall control
system is depicted in Figure 9.1. For the sake of simplicity the intermediate
variable s will be dropped as long as the omission causes no confusion.
It is well known that (9.1)∼(9.3) can equivalently be described by poly-
nomial matrix fraction representations:
Ay = B u + C d, (9.4)
Fξ= y, (9.5)
ym = E ξ, (9.6)
Pu= −Q d − R ym , (9.7)
where two pairs of polynomial matrices, A and [B C], and P and [−Q −R],
are both left coprime such that A−1 [B C] = [Gu Gd ] and P −1 [−Q −R] =
[−Gf − Gb ]; and F and E are right coprime such that EF −1 = Gm .
Combining (9.4)∼(9.7) yields
 9.1 Measurable Disturbance: Feedforward Control 295

    
−B 0 A u C
 0 F −I   ξ  =  0  d. (9.8)
P RE 0 y −Q
Let
 
−B 0 A
W =  0 F −I  , (9.9)
P RE 0
then the transfer function matrix between y and d is
Gyd = [0 0 I]W −1 [C T 0 − QT ]T . (9.10)
We can now formulate the disturbance decoupling problem with stability
as follows. (DDP): given a controlled system described by (9.1) and (9.2)
with Gu , Gd and Gm being proper rational matrices, find the conditions
under which a compensator in the form of (9.3) with Gf and Gb being proper
rational matrices will make (i) the resulting system stable, i.e., all the roots
of the determinant det(W ) lie in C− , the open left half of complex plane; and
(ii) the transfer matrix between y and d equals zero, i.e., Gyd = 0.

9.1.1 Solvability

Our strategy to resolve (DDP) is first to determine the set of proper compen-
sators which stabilize a controlled system. Within these stabilizable systems
and their compensators, the conditions for zeroing disturbance are then de-
rived.
Let G1 and G2 be rational matrices with the same number of rows, we
can factorize them as
G1 = D1−1 N1 , G2 = D2−1 N2 , (9.11)

[G1 G2 ] = D3−1 N3 , (9.12)

where all fractions are coprime. Let
D2 = D21 D22 , (9.13)
and partition N3 consistent with [G1 G2 ] as
N3 = [N31 N32 ]. (9.14)
Then the following can be established.

Lemma 9.1.1. With the above notations, D̃22 is a right divisor of any great-
−1
est common left divisor of N31 and D3 , where D̃22 D̃1 is a left coprime frac-
−1
tion of D1 D22 .
296 9. Dynamic Disturbance Decoupling

Proof. Equations (9.11), (9.12) and (9.14) show that G1 = D1−1 N1 =

D3−1 N31 . Furthermore, the coprimeness of N1 and D1 implies that [D3 N31 ] =
L1 [D1 N1 ] for some polynomial matrix L1 which is a greatest common left
divisor of N31 and D3 . Similarly, one can show that D3 = L2 D2 for some
polynomial matrix L2 . Grouping all these facts, we obtain
−1 −1
L2 = D3 D2−1 = L1 D1 D2−1 = L1 D1 D22 −1
D21 = L1 D̃22 −1
D̃1 D21 . (9.15)
−1
It follows from the coprimeness of D̃22 and D̃1 that L1 D̃22 must be a poly-
nomial matrix, that is, D̃22 is a right divisor of L1 . 
If D(s) is nonsingular, we will denote by σ(D) the set of all complex
numbers c such that det(D(c)) = 0.
Corollary 9.1.1. If σ(D̃22 ) is not empty, [D3 (c) N31 (c)] is of rank defect
for c ∈ σ(D̃22 ).
Proof. It follows from Lemma 9.1.1 that there is a polynomial matrix L3
such that L1 = L3 D̃22 and [D3 N31 ] = L3 D̃22 [D1 N1 ]. For c ∈ σ(D̃22 ),
[D3 (c) N31 (c)] is clearly of rank defect, and the corollary thus follows. 
Corollary 9.1.2. Let D̃2−1 D̃1 be a left coprime fraction of D1 D2−1 , then D3
and D̃2 D1 are left equivalent, i.e. there is a unimodular polynomial matrix U
such that D3 = U D̃2 D1 .
Proof. Set D22 = D2 and D21 = I. It follows from Lemma 9.1 that D3 =
L4 D̃2 D1 for some polynomial matrix L4 . One the other hand, we have
[G1 G2 ] = [D1−1 N1 D2−1 N2 ] = D1−1 [N1 D1 D2−1 N2 ].
= D1−1 [N1 D̃2−1 D̃1 N2 ] = (D̃2 D1 )−1 [D̃2 N1 D̃1 N2 ].
The last representation is also a fraction of [G1 G2 ] and can be related to
its coprime fraction D3−1 N3 by D̃2 D1 = L5 D3 for some polynomial matrix
L5 . Then, both L4 and L−14 = L5 are polynomial and it is unimodular. 
For a nonsingular polynomial matrix D, it can be factorized into D =
D− D+ such that σ(D−1 ) ⊂ C− and σ(D+ ) ⊂ C+ , the closed right half of
the complex plane (RHP). Such a factorization can be done, for example, by
determining the Smith form of D and then factorizing every invariant poly-
nomial as the product of a stable polynomial and an anti-stable polynomial.
For rational matrices G1 and G2 given above, this factorization is applied to
D2 , i.e., D2 = D2− D2+ . G1 is said to have the unstable model of G2 if D2+ is
a right divisor of D1 , i.e., D1 (D2+ )−1 is a polynomial matrix.
Theorem 9.1.1. Given a controlled system described by (9.1) and (9.2) with
Gu , Gd and Gm being proper, then a proper compensator described by (9.3)
stabilizes the system if and only if the following hold.
(i) Gu has the unstable model of Gd , and Gb the unstable model of Gf ,
(ii) Gm · Gu has no RHP pole-zero cancellations, and
(iii) Gb stabilizes Gm Gu .
 9.1 Measurable Disturbance: Feedforward Control 297

Proof. With the fraction representations given by (9.4)∼(9.7), we recall that

the compensated system is stable if and only if det(W ) has no zero in C+ .
An equivalent statement is
W (c) is of f ull rank f or any c ∈ C+ . (9.16)
Necessity: (i) Let Du−1 Nu and Dd−1 Nd be left coprime fractions of Gu
and Gd , respectively, and factorize Dd as Dd = Dd− Dd+ where σ(Dd− ) ⊂ C−
and σ(Dd+ ) ⊂ C+ . We now assume that Gu does not have the unstable
model of Gd . This means that Dd+ is not a right divisor of Du and σ(D̃d+ )
is not empty where (D̃d+ )−1 D̃u is a left coprime fraction of Du (Dd+ )−1 . Note
A−1 [B C] = [Gu Gd ] = [Du−1 Nu Dd−1 Nd ]. It follows from Corollary 9.1
that [A(c) B(c)] is of rank defect for c ∈ σ(D̃d+ ) ⊂ σ(Dd+ ) ⊂ C+ . Then W (c)
must be of rank defect for these c’s because
Rank[−B(c) 0 A(c)] = Rank[A(c) B(c)].
In a similar way, using
 
I 0 0
[P RE 0] = [P R 0]  0 E 0  ,
0 0 I
and
Rank[P (c) R(c)E(c) 0] ≤ Rank[P (c) R(c) 0],
one can show that if Gb does not have the unstable model of Gf , then W (c)
is of rank defect for some c ∈ C+ . These facts show that (i) is necessary for
(9.16) to hold.
(ii) We perform the following elementary row operation on W
    
IA0 −B 0 A −B AF 0
W1 = U1 W :=  0 I 0   0 F −I  =  0 F −I  . (9.17)
0 0 I P RE 0 P RE 0
Since U1 is unimodular, (9.16) is equivalent to the fact that W1 (c) is of full
rank for any c ∈ C+ . Then [−B AF 0] must thus be of full rank for any
c ∈ C+ , which implies
[A(c)F (c) B(c)] is of f ull rank f or any c ∈ C+ . (9.18)
In a similar way, postmultiplying W by the following unimodular matrix
 
I 0 0
U2 =  0 I 0,
0F I
one can show that (9.16) implies
 
A(c)F (c)
is of f ull rank f or any c ∈ C+ . (9.19)
E(c)
298 9. Dynamic Disturbance Decoupling

(9.18) and (9.19) together ensure that Gm · Gu = EF −1 A−1 B = E(AF )−1 B

has no RHP pole-zero cancellations.
(iii) From (9.17), it follows that
 
AF −B
det(W ) = det(W1 ) = a1 det
RE P

= a1 det(AF ) · det(P + RE(AF )−1 B), (9.20)

where a1 = +1 or −1. Using Corollary 9.1.2,
det(A) = a2 det(Du )det(D̃d ), (9.21)
where a2 is a nonzero constant and D̃d−1 D̃u is a left coprime fraction of
Du Dd−1 . Let
[P R] = L[P̄ R̄] (9.22)
for some polynomial matrix L such that P̄ and R̄ are coprime. Let
N D−1 = E(AF )−1 B, (9.23)
where N and D are coprime. Substituting (9.21) ∼ (9.23) into (9.20) yields
det(W ) = a · det(D̃d ) · det(L) · det(P̄ D + R̄N ) · g, (9.24)
where a is a nonzero constant, and g = det(Du ) · det(F )/det(D). Since
N D−1 = Gm Gu = EF −1 Du−1 Nu = E(Du F )−1 Nu , then g is a polynomial
and σ(g) is the set of all the cancelled poles when performing multiplica-
tion of Gm and Gu . It can be seen from (9.24) that for ensuring stability
det(P̄ D + R̄N ) must have no zeros in C+ , which also means that Gb stabi-
lizes Gm Gu . Necessity is thus proven.
Sufficiency: If Gu has the unstable model of Gd , then Du (Dd+ )−1 is a poly-
nomial matrix and thus σ(D̃d ) ⊂ σ(Dd− ) ⊂ C− . Similarly, one has σ(L) ⊂ C− .
Conditions (ii) and (iii) imply σ(g) ⊂ C− and σ(P̄ D + R̄N ) ⊂ C− , respec-
tively. It follows from (9.24) that σ(W ) ⊂ C− . Proof is therefore completed.


Remark 9.1.1. Theorem 9.1.1 is intuitively appealing. It indicates that in

order to ensure stability, all the unstable modes of Gd must be common to
Gu , i.e., Gd = (Dd+ )−1 G̃d and Gu = (Dd+ )−1 G̃u for some rational matrices
G̃d and G̃u , where G̃d is stable, and (Dd+ )−1 and G̃u have no RHP pole-
zero cancellations. There also hold Gf = (Df+ )−1 G̃f and Gb = (Df+ )−1 G̃b
for some rational matrices G̃f and G̃b , where G̃f is stable, and (Df+ )−1 and
G̃b have no RHP pole-zero cancellations. These show that all the unstable
modes (Dd+ )−1 and (Df+ )−1 of Gd and Gf must be within the feedback loop,
see Figure 9.2.
 9.1 Measurable Disturbance: Feedforward Control 299

d ~
Gd

~
Gf

- + y
~ u ~
Gb ( D +f ) −1 Gu ( Dd+ ) −1
- +

ym
Gm

Fig. 9.2. Possible Unstable Modes in Stabilizable Disturbance Decoupling Control

Remark 9.1.2. To know if Gu has the unstable model of Gd , one only needs
to check if Du Dd−1 is stable, which is much easier to perform without a need
to factorize Dd into Dd− Dd+ . It is also sufficient if Dd Gu is stable.
Consider now (DDP) under the stability conditions given in Theorem
9.1.1. From Figure 9.1, we have
Gyd = [I + Gu Gb Gm ]−1 [Gd − Gu Gf ].
Gyd = 0 is equivalent to
Gu Gf = Gd . (9.25)
Lemma 9.1.2. If the conditions of Theorem 9.1.1 are satisfied and Gf meets
(9.25), then Gf is stable.
Proof. Assume that Gf satisfies (9.25) but is not stable. With the notations
as in Theorem 9.1.1, premultiplying (9.25) by Dd+ gives
(D̄u )−1 Nu Gf = G̃d ,
where G̃d = (Dd− )−1 Nd and D̄u = Du (Dd+ )−1 with D̄u being a polynomial
matrix due to Gu having the unstable model of Gd . It follows from the sta-
bility of G̃d and coprimeness of Nu and D̄u that Nu · Gf must has RHP
pole-zero cancellations. Denote by (Df+ )−1 such a cancelled unstable model
of Gf . Then Gf = (Df+ )−1 G̃f , σ(Df+ ) ⊂ C+ , and Nu (Df+ )−1 is a poly-
nomial matrix. Since Gb has the unstable model of Gf , it must have the
form Gb = (Df+ )−1 G̃b . This results in the RHP pole-zero cancellations in
Gu · Gb = Du−1 Nu (Df+ )−1 G̃b , which violates the condition (ii) of Theorem
9.1.1. Proof is therefore completed. 
Theorem 9.1.2. (DDP) is solvable if and only if the following hold
(i) Gu has the unstable model of Gd ;
(ii) Gm · Gu has no RHP pole-zero cancellations; and
(iii) (9.25) has a proper stable solution Gf .
300 9. Dynamic Disturbance Decoupling

Proof. Assume that (DDP) is solvable, i.e., there are proper Gb and Gf such
that the system is stable and Gyd = 0. In view of Theorem 9.1.1, conditions
(i) and (ii) above must be satisfied in order to assure stability. In addition,
Gyd = 0 implies that Gf satisfies (9.25). It follows from Lemma 9.2 that Gf
must also be stable.
Conversely assume that the conditions (i) through (iii) are satisfied. Then
Gyd = 0 for the solution Gf . And there is a proper Gb which stabilizes Gm Gu .
Furthermore, such a Gb must have the unstable model of Gf due to Gf being
itself stable. Hence, by Theorem 9.1, the system is stable. And the above Gf
and Gb forms a solution of (DDP). 

The application of Theorem 9.1.2 to a special case yields the following

corollary.

Corollary 9.1.3. (DDP) using feedforward only (i.e., Gb = 0) is solvable if

and only if the following hold
(i) Gu , Gd and Gm are all stable; and
(ii) Equation (9.25) has a proper stable solution Gf .

9.1.2 Synthesis

Based on Theorem 9.1.2, a compensator which solves (DDP) can be con-

structed as follows.
Step 1. Check the solvability conditions. If they are satisfied, (DDP) is solv-
able and we can proceed to the next step; Otherwise, there is no
solution.
Step 2. Construct a feedforward compensator Gf . Any solution of (9.25) can
be taken as Gf provided that it is proper and stable. It is of course
desirable to choose its order as low as possible.
Step 3. Construct a feedback compensator Gb . A proper Gb is constructed
such that it stabilizes Gm Gu .
It is seen that Gb and Gf can be constructed independently. In order to check
whether (i) and (ii) of Theorem 9.1.2 are satisfied, only matrix factorizations
and divisions are needed and they have been well covered in Chapter 2.
Stabilization problem in Step 3 has been discussed in Chapter 3. Therefore,
our attention here will be paid only to (iii) of Theorem 9.1.2 and Step 2
above so as to find when (9.25) has a proper stable solution and how it can
be calculated.
Given a m × l rational matrix Ḡ1 of rank r and a m × p rational matrix
Ḡ2 , an l × p proper rational matrix G needs to be found such that
Ḡ1 G = Ḡ2 . (9.26)
This is called the model-matching problem. We now present a unified ap-
proach, which can treat the properness, stability and minimality simultane-
ously, and give an explicit parameterization of all the solutions.
 9.1 Measurable Disturbance: Feedforward Control 301

It is well known that there is a unimodular polynomial matrix U such

that
 
G1
U Ḡ1 = ,
0
where G1 has full row rank r. And the same operation is applied to Ḡ2 :
 
G2
U Ḡ2 = 0 ,
G2
where G2 has the same number of rows as G1 . (9.26) is clearly equivalent to
0
G2 = 0, (9.27)
G1 G = G2 . (9.28)
Let a polynomial matrix X be such that X[G1 − G2 ] := P is a polynomial
matrix. Elementary row operations are performed on P to make P row re-
duced. This means that there is an unimodular polynomial matrix U1 such
that U1 P := P̃ is row reduced. P̃ is partitioned as
P̃ = [P̃1 − P̃2 ], (9.29)
which is consistent with [G1 − G2 ]. A greatest common left divisor of the
columns of P1 is denoted by L, i.e., P̃1 = LP̄1 , where columns of polynomial
matrix P̄1 are left coprime. Let B be a minimal base of the right null space
of P̃1 and P̄2 R−1 be a dual coprime fraction of L−1 P̃2 . Denote by ∂ci (A)
the i th column degree of a polynomial matrix A, and by Γr (A) the highest
row-degree coefficient matrix of Ã. Γr (P̃ ) is partitioned as
Γr (P̃ ) = [P̃h1 P̃h2 ],
which is consistent with (9.29).

Theorem 9.1.3. Given Ḡ1 and Ḡ2 as above, (9.26) has a solution for G if
and only if (9.27) holds. If it is the case, all the solutions are given by the
following polynomial matrix fraction:
G = N D−1 , (9.30)
N = BW + N0 V, (9.31)
D = RV, (9.32)
where W and V are arbitrary polynomial matrices, and N0 satisfies
P̃1 N0 = P̃2 R. (9.33)
Furthermore,
(i) a solution is of minimal order if and only if V is unimodular. Without
loss of generality all the minimal solutions are given by (9.30) and (9.33)
with V = I;
302 9. Dynamic Disturbance Decoupling

(ii) There is a proper solution if and only if

Rank[P̃h1 ] = r. (9.34)
If it holds, all the proper solutions are given by (9.30) through (9.33)
with N0 R−1 being proper, D column reduced and ∂ci (BW ) ≤ ∂ci (D);
And
(iii) there is a stable solution if and only if det(R) is a stable polynomial. If
it is, all the stable solutions are given by (9.30) through (9.33) with V −1
being stable.

Proof. It is clear that (9.27) is necessary for the existence of solutions. Con-
versely assume that it holds. Then (9.26) will have a solution if (9.28) has a
solution, which is guaranteed due to G1 having full rank. Indeed, replace G
in (9.28) by a right coprime polynomial matrix fraction,
G = N D−1 . (9.35)
Equation (9.28) can be rewritten as
P̃1 N = P̃2 D. (9.36)
−1
Premultiplying (9.36) by L yields
P̄1 N = P̄2 R−1 D. (9.37)
Because the columns of P̄1 are coprime by the definition of L, then there is
a polynomial matrix M such that P̄1 M = I. Setting N = M P̄2 and D = R,
then such N and D satisfy (9.37) and thus N D−1 is a solution of (9.28).
In order to search for a general solution of (9.36), it will be first shown
that if D is any solution of (9.36), it must have the form
D = RV, (9.38)
for some polynomial matrix V . Indeed, assume that D is a solution of (9.36),
it then satisfies (9.37). This implies that R−1 D must be a polynomial matrix
because P̄2 and R are coprime. In other words, D = RV for some polynomial
matrix V .
A general solution of (9.36) can be expressed as the sum of the general
solution of the associated homogeneous equation,
P̃1 N1 = 0, (9.39)
and a particular solution N2 of the inhomogeneous equation (9.36). If B is a
minimal base of the right null space of P̃1 , then it is well known (Forney 1975)
that the general solution N1 of (9.39) is N1 = BW , where W is a arbitrary
polynomial matrix. For the particular solution N2 of (9.36), set N2 = N0 V ,
where N0 satisfies
P̃1 N0 = P̃2 R, (9.40)
 9.1 Measurable Disturbance: Feedforward Control 303

then N2 is clearly a solution of (9.36) due to D = RV as shown above.

Therefore, we have
N = N1 + N2 = BW + N0 V. (9.41)
Next, we prove (i)-(iii).
(i) The order of a solution is deg det(D), where deg det(D) = deg det(R)+
deg det(V ) ≥ deg det(R). The minimal order is achieved if and only if
deg det(V ) = 0, i.e., V is unimodular. It is clear that V can be taken as
I, which does not change the form of (9.30) if W is replaced by W V −1 . And
W V −1 is equivalent to W since V is unimodular.
(ii) It has been shown (Kung and Kailath, 1980) that there is a proper
minimal solution if and only if (9.34) hold. If it does, there is a proper minimal
solution which is, of course, proper. Conversely if a proper solution exists,
then in the set of all the proper solutions, there must be one with minimal
order. Thus (9.34) is necessary and sufficient for the existence of a proper
solution. Some constraints will be imposed on the general solution given by
(9.30) to (9.33) in order to obtain all the proper solutions. For the simplicity,
assuming D is column reduced. It can be easily seen that this is not a re-
striction since there is always a unimodular polynomial matrix U such that
DU is column reduced. If N and D is replaced by N U and DU , respectively,
no solution will be changed or lost. Therefore D can be assumed column
reduced. As mentioned above, when (9.34) holds, there is a proper minimal
solution of (9.36). But one sees from the proof of (i) that such a solution can
be obtained by setting V = I. Now let N0 be a solution of (9.33) such that
N0 R−1 is proper. This can easily be done by any algorithm for the minimal
design problem. For such a N0 , all the solutions of (9.28), as shown above,
can be given by (9.30). Furthermore, since N D−1 = (BW )D−1 + N0 R−1
with N0 R−1 being proper, then the solutions will be proper if and only if
(BW )D−1 is proper, which is equivalent to ∂ci (BW ) ≤ ∂ci (D) when D is
column reduced.
(iii) Because R is common to all the solutions and V is arbitrary, there is
a stable solution if and only if det(R) is a stable polynomial. If it is the case,
all the stable solutions are given by (9.30) to (9.33) with V −1 being stable.


For illustration, consider

" 1 1 #
s−1 (s−1)(s+2)
Gu = 1 1
,
s−1 (s−1)(s+1)
" 1 #
(s−1)(s+2) 0
Gd = 1 1
,
(s−1)(s+1) (s−1)(s+1)(s+2)
Gm = I.
Clearly Gd has an unstable pole at s = 1. We got left coprime fractions
Gu = Du−1 Nu and Gd = Dd−1 Nd , where
304 9. Dynamic Disturbance Decoupling

   
(s − 1)(s + 2) 0 (s + 2) 1
Du = , Nu = ,
0 (s − 1)(s + 1) (s + 1) 1
   
(s − 1)(s + 2) 0 1 0
Dd = , Nd = .
0 (s − 1)(s + 1)(s + 2) (s + 2) 1
Because
" #
1 0
Du Dd−1 = 1
0 s+2

is stable so that Gu has the unstable model of Gd . For Gm = I, Gm · Gu has

no RHP pole-zero cancellations. It follows from Theorem 9.1.2 that there is
a solution for (DDP) if and only if the equation,
Gu G f = G d ,
has a proper stable solution. According to Theorem 9.1.3, it has a solution
due to nonsingularity of Gu . Premultiplying [Gu − Gd ] by Dd yields
 
  (s + 2) 1 −1 0
P = Dd Gu −Gd =
(s + 1)(s + 2) (s + 2) −(s + 2) −1
which is left equivalent to
 
(s + 2) 1 −1 0
P̃ =
0 1 −1 −1
which is row reduced. One sees
 
1 0 0 0
Γr (P̃ ) = ,
0 1 −1 −1

Rank(P̃h1 ) = 2.
Thus a proper solution exists. Because P̃1 = P̃1 · I, the greatest common left
divisor of P̃1 can be taken as itself, i.e., L = P̃1 . And the minimal base B of
L is clearly zero. Hence
 −1  
−1 (s + 2) 1 10
L P̃2 =
0 1 11
" 1
#   −1
0 − s+2 0 −1 1 0
= = := P̄2 R−1 . (9.42)
1 1 1 s + 2 0 s + 2

det(R) = (s + 2) is a stable polynomial so that a stable solution exists.

Therefore, (DDP) for this example is solvable. It is now possible to give an
explicit parametric form of all the proper stable solutions. In order to search
for a particular solution N0 of the equation: P̃1 N = P̃2 R, for N such that
N0 R−1 is proper. Set N0 = P̄2 which is clearly a needed solution. A proper
and stable solution is then given by Gf = P̄2 R−1 as given by (9.42).
It can be noted that for this example, since Gu is nonsingular, Gf is in
fact unique and can be obtained directly by Gf = G−1 u Gd . In this case (DDP)
 9.2 Unmeasurable Disturbances: Disturbance Observers 305

is solvable if and only if this Gf is proper and stable. The computations in the
example is thus not necessary, but is carried out to illustrate the procedure
for a general case.
For the feedback compensator Gb which stabilizes Gu there are many
well-known algorithms for its construction (Chapter 3). Therefore, specific
computations are omitted here.
In this section the disturbance decoupling problem with stability by dy-
namic measurement feedback and disturbance feedforward has been consid-
ered. The problem is solved in a polynomial matrix setting. The stability
conditions are established. The main observation is that for ensuring the sta-
bility, the transfer matrix of control channel and the feedback compensator
must have the unstable models of disturbance channel and of the feed-forward
compensator, respectively, and there should be no RHP pole-zero cancella-
tions between the plant and sensor. They are intuitively appealing and have
an explicit dynamic interpretation. Under such stability conditions it is then
shown that solvability of disturbance decoupling is equivalent to the exis-
tence of a proper stable solution for some rational matrix equation. It is also
shown that the construction of a compensator for (DDP) can be made into
two independent parts. One is the feedback compensator synthesis, which is
a well-known problem. Another is the feed-forward compensator synthesis,
which can be incorporated into a more general problem, namely the model
matching. A unified approach, which can simultaneously treat properness,
stability and minimality, is presented. This approach also gives an explicit
parametric form of all the solutions.

9.2 Unmeasurable Disturbances: Disturbance Observers

In this section, supposing unmeasurable disturbance, we present a general

scheme for disturbance decoupling control via output feedback for processes
with possibly time delay. The objective is to fully compensate for the dis-
turbance while retaining the input-output servo response of the conventional
feedback system. Internal stability and disturbance rejection performance will
be analyzed, and the special case of periodic disturbances highlighted. SISO
systems will be treated first and then extended to the MIMO systems.
Our development will involve both time and Laplace domain representa-
tions of signals. For clarity, the former is denoted by small case letters such
as d(t) while the latter by capitals such as D(s). Throughout this section,
we assume that the process is stable and has no zeros in the right half plane
(RHP), unless otherwise stated. The diagram of the conventional feedback
control system is depicted in Figure 9.3, where P (s) and C(s) denote the pro-
cess and feedback controller respectively, and d(t) is an unknown disturbance
acting on the system through a stable dynamic element G(s). It follows from
Figure 9.3 that
306 9. Dynamic Disturbance Decoupling

PC G
Yo = R+ D,
1 + PC 1 + PC
where the superscription o indicates the output response in the conventional
feedback system.

Fig. 9.3. Conventional Feedback System

9.2.1 Disturbance Observers

A control scheme, depicted in Figure 9.4, is exploited to compensate for the

disturbance. The intention is to estimate the disturbance by the difference
between the actual process output and its model output, and feed the esti-
mate into another block to fully cancel the disturbance. From Figure 9.4, one
has
PC 1 − HF
Y = R+ GD. (9.43)
1 − HF + P C + P H 1 − HF + P C + P H
To fully compensate for the disturbance while retaining the input-output
relation for servo response of the conventional feedback system requires
PC PC
= ,
1 − HF + P C + P H 1 + PC
1 − HF
GD = 0, (9.44)
1 − HF + P C + P H
which are satisfied by

F (s) = P (s), (9.45)

 9.2 Unmeasurable Disturbances: Disturbance Observers 307

Fig. 9.4. Control Scheme with Disturbance Observer

and
H(s) = P −1 (s). (9.46)

Let P (s) = P0 (s)e−Ls and its model P̂ (s) = P̂0 (s)e−L̂s , where P0 and P̂0
are delay-free rational transfer functions. Then, using the model, (9.45) and
(9.46) are replaced by

F (s) = P̂ (s) = P̂0 (s)e−L̂s , (9.47)

1 1
H(s) = = eL̂s , (9.48)
F (s) P̂0 (s)

respectively. For a strictly proper P̂0 (s), its inverse is not physically realiz-
able, and nor is the pure predictor eL̂s . For the term eL̂s , the approximation
presented in Huang et al. (1990) as shown in Figure 9.5 is adopted here,
where the overall transfer function is
1 + V (s)
Gv (s) = .
1 + V (s)e−L̂s
If V (s) is of low-pass with high gain, then
V (s)
Gv (s) ≈ = eL̂s
V (s)e−L̂s
308 9. Dynamic Disturbance Decoupling

in the low-frequency range, while Gv (s) ≈ 1+0 1+0 = 1 in the high-frequency

range. The simplest choice for V (s) is a first-order element:
Kv
V (s) = , τv > 0,
τv s + 1
and then the corresponding Gv is given by
Kv
1+ τv s+1
Gv (s) = Kv
. (9.49)
1+ −L̂s
τv s+1 e

According to our study, the value of τv is suggested to be set as 0.1L̂ ∼ L̂. The
remaining parameter Kv is chosen as large as possible for high low-frequency
gain of Gv while preserving stability of Gv . The critical value Kvmax of Kv
at which the Nyquist curve of τvKs+1
v
e−L̂s touches −1 + j0 can be found from
q
Kvmax = 1 + ω ∗2 τv2 ,
tan−1 (ω ∗ τv ) + ω ∗ L̂ = π. (9.50)
For sufficient stability margin, Kv should be much smaller than Kvmax and
its default value is suggested to be Kvmax /3.
P̂0 (s) is usually strictly proper and P̂ 1(s) is then improper. A filter Q(s)
0
Q(s)
is introduced to make P̂0 (s)
proper and thus realizable. Q(s) should have at
least the same relative degree as that of P̂0 (s) to make Q(s)/P̂0 (s) proper,
and should be approximately 1 in the disturbance frequency range for a sat-
isfactory disturbance rejection performance. The following form is a possible
choice:
1
Q(s) = , τq > 0, (9.51)
(τq s + 1)n

where n is selected to make P̂Q(s) just proper. τq is a user specified parameter

0 (s)
and should be smaller than the equivalent process time constant Tp . The

Fig. 9.5. Block Diagram of Gv (s) Approximating eL̂s

 9.2 Unmeasurable Disturbances: Disturbance Observers 309

initial value is set as 0.01Tp ∼ 0.1Tp and its tuning will be discussed below to
make the system stable. Other forms of Q(s) have been discussed by Umeno
and Hori (1991).
In view of the above development, H(s) is approximated by

Q(s)Gv (s) Q(s) 1 + V (s)

H(s) = = . (9.52)
P̂0 (s) P̂0 (s) 1 + V (s)e−L̂s

The actual system to realize the proposed control scheme with such an ap-
proximation is depicted in Figure 9.6.

G (s )

r + + u + + y
C (s ) P0 ( s ) e − Ls
_ _

+
Pˆ0 ( s ) ˆ
e − Ls _

+
Q( s )
V (s )
+ _ + Pˆ ( s )
0

ˆ
e − Ls
Gv (s )

Fig. 9.6. Control Scheme with Approximate Disturbance Observer

If P̂ (s) is bi-proper (the degrees of the numerator and denominator are

the same), the filter Q(s) is not required. For a dead-time free case, Gv is not
required or Gv (s) = 1, (9.52) becomes

Q(s)
H(s) = ,
P̂ (s)

which is the original disturbance observer proposed by Ohnishi (1987).

310 9. Dynamic Disturbance Decoupling

9.2.2 Analysis

Consider stability and performance of the control scheme in Figure 9.6. Let
a system consist of n subsystems with scalar transfer functions gi (s), i =
1, 2, · · · , n, and pi (s) be the characteristic polynomial of gi (s). Define
pc (s) = ∆(s)p0 (s), (9.53)
where ∆(s) is the system Q determinant as given by the Mason’s formula
n
(Mason, 1956) and p0 (s) = i=1 pi (s). It is shown (Chapter 3) that the sys-
tem is internally stable if pc (s) in (9.53) has all its roots in the open left-half
of the complex plane.
For the proposed scheme, write the transfer function of each block into
a coprime polynomial fraction with possible time delay, i.e., C(s) = ndcc(s) (s)
,
np (s) −Ls np̂ (s) −L̂s nv (s) nq (s)
P (s) = dp (s) e , P̂ (s) = dp̂ (s) e , V (s) = dv (s) , Q(s) = dq (s) and G(s) =
ng (s)
dg (s) ,
where nc (s), dc (s), np (s), dp (s), np̂ (s), dp̂ (s), nv (s), dv (s), nq (s) dq (s),
ng (s) and dg (s) are all polynomials. It follows from Figure 9.6 that
P0 P0
∆(s) = 1 + P C + QV e−Ls + Qe−Ls
P̂0 P̂0
− QV e−L̂s − Qe−L̂s + V e−L̂s + P CV e−L̂s ,
P0
= (1 + P C)(1 + V e−L̂s ) + (1 + V )Q( e−Ls − e−L̂s ), (9.54)
P̂0
p0 (s) = np̂ (s)dp (s)dp̂ (s)dv (s)dc (s)dq (s)dg (s). (9.55)

Under the nominal case, i.e., P̂ (s) = P (s), ∆ becomes

∆ˆ = (1 + P C)(1 + V e−Ls ),
so that
 þ
ˆ 0 = {dp dc (1 + P C)} dv (1 + V e−Ls ) {np dp dq dg } .
p̂c = ∆p
The assumed stability of the conventional feedback system in Figure 9.3
means that dp dc (1 + P C) only has LHP roots. Under the assumptions that
the process is stable and without RHP zero, and G is stable, dp , dg and np
are then all stable. Therefore, it is concluded that the proposed scheme is
nominally stable, provided that stable Gv (equivalently dv (1 + V e−Ls )) and
Q (or dq ) are used.
Turn now to the case where there is a process/model mismatch:

P̂ (s) 6= P (s),
and the modelling error is described by
ü ü
ü ü
üP (jω) − P̂ (jω)ü < γ(jω).
 9.2 Unmeasurable Disturbances: Disturbance Observers 311

Then, pc (s) can be re-written as

pc (s) ˆ + δ(s),
= ∆(s) = ∆(s)
p0 (s)

where the perturbed term from the nominal case is δ(s) = (1+V )Q( P
P̂
0 −Ls
e −
0

e−L̂s ) and
ü ü
ü 1 + V (jω) ü
ü ü
|δ(jω)| = ü Q(jω)(P (jω) − P̂ (jω))ü
ü P̂0 (jω) ü
ü ü
ü 1 + V (jω) ü
ü ü
<ü Q(jω)ü γ(jω) := δ̄(jω).
ü P̂0 (jω) ü

Since the nominal stability has been proven, the Nyquist curve of ∆(s) ˆ
should have a correct number of encirclements of the origin, as required by
ˆ
the Nyquist stability criterion. One can superimpose the δ̄(jω) onto ∆(s)’s
Nyquist curve to form a Nyquist curve band, and check its stability.
Consider now the performance of the proposed method and its robustness.
With a stabilizing and tracking controller C, the process output response
yd (t) to a disturbance will settle down, i.e., yd (t) → 0 as t → ∞. Thus, the
widely used index in the time domain:
sZ
∞
kyd (t)k2 = yd2 (t)dt,
0

makes sense and can be used for performance evaluation. To have a better
disturbance rejection than the conventional feedback system requires
kyd (t)k2 6 kydo (t)k2 , (9.56)
where ydo (t) is the output response to the same disturbance if the conventional
feedback system is used. Applying the Parseval’s Identity (Spiegel, 1965):
Z ∞ Z ∞
1
yd2 (t)dt = |Yd (jω)|2 dω,
0 2π 0
to (9.56) gives
kYd k2 6 kYdo k2 , (9.57)
where
sZ
∞
kYd k2 , |Yd (jω)|2 dω,
0
sZ
∞
kYdo k2 , |Ydo (jω)|2 dω.
0
312 9. Dynamic Disturbance Decoupling

Substituting (9.43), (9.47) and (9.52) into (9.57) yields

û û û û
û 1 − Gv Qe−L̂s û û GD û
û û û û .
û GDû 6 û (9.58)
û 1 + P C + Gv Q( P0 e−Ls − e−L̂s ) û 1 + P C û2
P̂0 2

In case of P̂ (s) = P (s), (9.58) becomes

û û û û
û GD û û GD û
û −Ls û û û .
û 1 + P C (1 − Gv Qe )û 6 û
1 + P C û2
(9.59)
2

e−Ls
For illustration, let us consider the simple case with Gv = eLs , P = Tp s+1
and C(s) = kc + ksi . The disturbance is supposed to be D(s) = 1
s through a
1 1
dynamics G(s) = Td s+1 , and the filter is supposed to be Q(s) = τq s+1 . With
simple algebra, we evaluate
û û Z ∞
û GD û2
û û = A(ω)dω, (9.60)
û1 + PC û
2 0

where
A(ω)
Tp2 ω 2 + 1
=
(Td2 ω 2 + 1)[(−Tp ω 2 + ki cos ωL + kc ω sin ωL)2 + (ω + kc ω cos ωL − ki sin ωL)2 ]
> 0.

Similarly, the left-hand side of (9.59) becomes

û û2 Z ∞
û GD û τq2 ω 2
û (1 − Q) û = A(ω) dω (9.61)
û1 + PC û τq2 ω 2 + 1
2 0
Z ∞ û û
û GD û2
6 A(ω)dω = û û û , (9.62)
0 1 + P C û2
so that (9.59) holds. The performance enhancement of the proposed method
over the conventional feedback system is ensured for this special case in the
sense of (9.56).

9.2.3 Periodic Disturbances

In the discussion made so far, no specific characteristic is assumed for the

disturbance. If the disturbance is of a special type and this information is
utilized beneficially, disturbance rejection performance may be further en-
hanced. In this subsection, a special type of disturbances, periodic distur-
bances, will be investigated in details. Periodic disturbances are often en-
countered in mechanical systems such as industrial robots (Hara et al., 1988).
Disturbances acting on the track-following servo systems of an optical disk
 9.2 Unmeasurable Disturbances: Disturbance Observers 313

drive or a hard disk drive inherently contain significant periodic components

that cause tracking errors of a periodic nature. Also, continuous periodic
patterns/disturbances are abundant in chemical process industries and ex-
amples include pressure swing absorption systems extruders, hybrid reactors
and simulated moving bed (SMB) chromatography systems (Natarajan and
Lee, 2000).
We begin with some definitions and preliminary results. Let a periodic
signal x(t) have a period T and act as an input to the stable system. The
output steady state response will also be periodic with the same period. Let
the periodic steady state part of the output response y(t) be y s (t), and the
corresponding Laplace transform be Y s (s) = L{y s (t)}.
Lemma 9.2.1. Under a periodic input x(t), the output steady state of the
system Y (s) = G2 (s)G1 (s)X(s) with stable G2 (s) and G1 (s) satisfies

(g2 ∗ g1 ∗ x)s (t) = [g2 ∗ (g1 ∗ x)s ]s (t),

where ∗ is the standard convolution, and g1 and g2 are the impulse responses
of G1 (s) and G2 (s) respectively.

Proof. Suppose that x(t) has the period T and is expressed by its Fourier
series (Cartwright, 1990):
∞
X
x(t) = x0 + Ci cos(2πit/T + φi )
i=1
X∞
= x0 + Ci cos(ωi t + φi )
i=1
RT p
with x0 = T1 0 xdt, Ci = A2i + Bi2 , ωi = 2πi/T , and φi = − tan−1 (Bi /Ai ),
RT RT
where Ai = T2 0 cos(2πit/T )xdt and Bi = T2 0 sin(2πit/T )xdt. Having
such a signal transmitted through a stable G1 , the output is given by g1 (t) ∗
x(t). The steady state of the output is a periodic signal with the same period
T (Cartwright, 1990):
∞
X
(g1 ∗ x)s (t) = G1 (0)x0 + Ci |G1 (jωi )| cos(ωi t + φi + ∠G1 (jωi )).
i=1

Similarly, one gets

∞
X
[g2 ∗ (g1 ∗ x)s ]s (t) = G2 (0)G1 (0)x0 + Ci |G1 (jωi )||G2 (jωi )|
i=1
cos(ωi t + φi + ∠G1 (jωi ) + ∠G2 (jωi )).

On the other hand, it follows from

314 9. Dynamic Disturbance Decoupling

g2 (t) ∗ g1 (t) ∗ x(t) = [g2 (t) ∗ g1 (t)] ∗ x(t)

that y s (t) is directly given by

∞
X
(g2 ∗ g1 ∗ x)s (t) = G2 (0)G1 (0)x0 + Ci |G2 (jωi )G1 (jωi )|
i=1
cos(ωi t + φi + ∠G2 (jωi )G1 (jωi )).

Noting that

|G1 (jωi )G2 (jωi )| = |G2 (jωi )||G1 (jωi )|,

and

∠[G2 (jωi )G1 (jωi )] = ∠G1 (jωi ) + ∠G2 (jωi ),

we then have

y s (t) = (g2 ∗ g1 ∗ x)s (t) = [g2 ∗ (g1 ∗ x)s ]s (t). 

Turn back to the rejection of periodic disturbances, we use the same F

as in (9.47) but for H replace (9.48) by
1
H(s) = e−Lh s , (9.63)
P̂0 (s)

where Lh = k0 T − L̂, and k0 is the smallest no-negative integer that renders

a non-negative Lh > 0. With the preceding discussions, Figure 9.6 is recasted
as Figure 9.7. Substituting (9.63) into (9.43) yields

1 − e−k0 T s
Yd (s) = GD. (9.64)
1 − HF + P C + P H
Let η(t) the impulse response of G/(1 − HF + P C + P H). One obtains from
(9.64) that

yds (t) = [η(t) ∗ L−1 {(D(s) − e−k0 T s D(s))}]s

= (η(t) ∗ (d(t) − d(t − k0 T )))s
= (η(t) ∗ (ds (t) − ds (t − k0 T )))s
= (η(t) ∗ 0)s
= 0. (9.65)

Thus, the periodic disturbance can be asymptotically rejected. For physi-

cal realization of H(s) in (9.63), a filter Q, as discussed above, has to be
introduced to make Q(s)/P̂0 (s) proper.
 9.2 Unmeasurable Disturbances: Disturbance Observers 315

G (s )

r + + u + + y
C (s ) P0 ( s ) e − Ls
_ _

+
Pˆ0 ( s ) ˆ
e − Ls _

Q( s )
e − Lh s
Pˆ0 ( s )

Fig. 9.7. Control Scheme for Periodic Disturbance Rejection

9.2.4 Simulation

We now present several examples to demonstrate the disturbance observer

based control and compare with conventional feedback systems.
Example 9.2.1. Consider a first-order plus dead-time process (Astrom and
Hagglund, 1995):

e−5s
P (s) = .
10s + 1
The PI controller design method proposed in Zhuang and Atherton (1993) is
applied to give
1.01
C(s) = 1.01 + .
s
For the scheme in Figure 9.6 with perfect model assumption P̂ (s) = P (s),
take τv = 0.4L = 2. We obtain Kvmax = 1.38 from (9.50). Setting Kv =
0.46
Kvmax /3 = 0.46 yields V (s) = 2s+1 . The filter Q(s) is in the form of Q(s) =
1
τq s+1 , and τq is chosen as 0.1Tp = 1. A unit step change in the set-point is
made at t = 0, and a step disturbance of the size 1 is introduced through
1
G(s) = 20s+1 to the system at t = 50, the resulting responses are displayed in
Figure 9.8. Clearly, the disturbance observer control exhibits better capability
to reject load disturbance compared with the conventional controller.
To evaluate the scheme of Figure 9.7 for periodic disturbances, the fol-
lowing periodic disturbance is introduced at t = 50 as a separate test:
316 9. Dynamic Disturbance Decoupling

1
d

0.5

0
0 10 20 30 40 50 60 70 80 90 100

1.5

1
y

0.5

0
0 10 20 30 40 50 60 70 80 90 100

1
u

−1
0 10 20 30 40 50 60 70 80 90 100

Fig. 9.8. Time Responses for Example 9.2.1

(—— Disturbance Observer; − − − Conventional Control )

d(t) = sin(0.2πt) + 0.3 sin(0.4πt) + 0.1 sin(0.6πt) + 0.1 sin(0.8πt)

2π
with the period of T = 0.1π = 10. Let G(s) = 1. Calculate from (9.63)
Lh = T − L = 5. In the case of no plant/model mismatch, the resultant
responses are shown in Figure 9.9. The output response of the proposed
method approaches zero asymptotically, as expected from (9.65). Comparing
with the conventional control, the disturbance observer control produces great
improvement in the load response. ♦
Example 9.2.2. Consider a process with multiple leg:
1
P (s) = ,
(s + 1)8
with its model (Bi et al., 1999):
1.06
P̂ (s) = e−4.94s .
3.81s + 1
A PI controller is designed (Zhuang and Atherton, 1993) as
0.0967
C(s) = 0.3685 + .
s
 9.2 Unmeasurable Disturbances: Disturbance Observers 317

1
0.5
0
d

−0.5
−1
0 10 20 30 40 50 60 70 80 90 100

1
y

−1
0 10 20 30 40 50 60 70 80 90 100

10

0
u

−10

−20
0 10 20 30 40 50 60 70 80 90 100

Fig. 9.9. Time Responses for Example 9.2.1 under Periodic Disturbance
(——Disturbance Observer; − − −Conventional Control)

For the scheme of Figure 9.6, V (s) = 0.4625 1

2s+1 and Q(s) = 5s+1 are designed.
Using P and P̂ in the scheme, the system step responses are shown in Fig-
ure 9.10, and improvement in the load response is substantial with the dis-
turbance observer control. To verify stability, we draw the Nyquist curve
ˆ
of ∆(s) with the perturbation term δ(s) superimposed onto it, as displayed
in Figure 9.11. The resulting banks neither enclose nor encircle the origin.
Therefore, the system with the actual process P in place is indeed stable. ♦

In the development made so far, the process P (s) has been assumed to
be stable without any zeros in the RHP. However, the control scheme in
Figure 9.6 can be modified to Figure 9.12 so that it is applicable to unstable
processes and/or those with RHP zeros. In this case, the process P (s) is
factorized as

P (s) = P0+ (s)P0− (s)e−Ls ,

where P0− (s) represents the stable non-minimum-phase part, and P0+ (s) in-
cludes all RHP zeros and poles. Subsequently, P0− (s) is used in (9.47) and
(9.52) to design F (s) and H(s), respectively.
318 9. Dynamic Disturbance Decoupling

1
d

0.5

0
0 50 100 150

1.5

1
y

0.5

0
0 50 100 150

1.5

1
u

0.5

0
0 50 100 150

Fig. 9.10. Time Responses for Example 9.2.2

(—— Disturbance Observer; − − − Conventional Control)

Example 9.2.3. Consider an unstable process:

e−0.4s
P (s) = .
(s + 1)(s − 1)

A PID controller is designed (Ho and Xu, 1998) as

0.1175
C(s) = 1.8075 + + 1.6900s.
s
For the scheme of Figure 9.12 with the modification mentioned above, V (s) =
0.3962 1
0.1s+1 and Q(s) = 10s+1 are designed. The resulting performance is shown
in Figure 9.13, where the effectiveness of the disturbance observer control is
clearly illustrated. ♦

9.2.5 Extension to the MIMO Case

We have so far considered the SISO case. The extension to the MIMO case is
straightforward. Suppose that the process P (s) is square, nonsingular, stable
and of minimum phase. The same scheme as in Figure 9.4 is exploited. It
then follows that
 9.2 Unmeasurable Disturbances: Disturbance Observers 319

Im
1

−1

−2

−3

−4

−5

−6

−7
Re
−0.2 0 0.2 0.4 0.6 0.8 1 1.2

ˆ − − − δ)
Fig. 9.11. Nyquist Curve Banks For Example 9.2.2 (—— ∆;

Y = GD + P U, (9.66)

and
U = C(R − Y ) − H(Y − F U )
= CR + HF U − (C + H)(GD + P U ),

which is then solved for U . Substituting this U into (9.66) gives

Y = P [I − HF + CP + HP ]−1 CR + [I − P [I − HF + CP + HP ]−1 (C + H)]GD.
(9.67)

Use F and H as in (9.45) and (9.46), respectively, that is,

F (s) = P (s), H(s) = P −1 (s).

Equation (9.67) becomes

Y = P [I + CP ]−1 CR + [I − P [I + CP ]−1 (C + P −1 )]GD
= [I + P C]−1 P CR + [I − [I + P C]−1 P (C + P −1 )]GD
= [I + P C]−1 P CR, (9.68)
which is independent of the disturbance D. For implementation of F and H,
we use a process model P̂ in place of P as usual. Now the inversion of P̂ and
its realization need our attention. Suppose also that P̂ is nonsingular. Inverse
320 9. Dynamic Disturbance Decoupling

Fig. 9.12. Disturbance Observer Scheme for Unstable/non-minimum-phase Pro-

cesses

P̂ (jω) to get P̂ −1 (jω). Apply the model reduction (Chapter 2) to P̂ −1 (jω)

to get its approximation P̃ (s) in form of

p̃ij (s) = p̃ij0 (s)e−Lij s . (9.69)

where p̃ij0 (s) are proper rational functions. But terms e−Lij s often happen to
be a pure predictor which is not physically realizable, that is, some Lij may be
negative. They will require the approximation shown in Figure 9.5, each for
one negative Lij . It is costly. To reduce the number of such approximations,
extract the largest possible pure prediction horizon L by

0, if all − Lij ≤ 0;
L= (9.70)
max{−L | − L > 0}, otherwise.
ij ij

Write −Lij = L − (L + Lij ) with L∗ := L + Lij ≥ 0 due to L ≥ −Lij so that

∗
e−Lij s correspond to time delays. Equation (9.69) can be rewritten as
 9.2 Unmeasurable Disturbances: Disturbance Observers 321

1.05

0.95

0.9
y

0.85

0.8

0.75

0.7
0 50 100 150

0.5

−0.5
u

−1

−1.5
0 50 100 150
−0.4s
e
Fig. 9.13. Time responses for P (s) = (s+1)(s−1)
(—— Disturbance Observer; − − − Conventional Control)

 ∗ ∗ ∗ 
p̃110 (s)e−L11 s p̃120 (s)e−L12 s · · · p̃1m0 (s)e−L1m s
 
 
 
P̃ = diag {L, L, · · · , L} ··· ··· ··· ··· 
 
 
∗ ∗ ∗
p̃m10 (s)e−Lm1 s p̃m20 (s)e−Lm2 s · · · p̃mm0 (s)e−Lmm s
(9.71)

so that we only need to have m approximations to pure predictors instead of

possible m × m ones, for m × m processes. Besides, the design involving the
approximation in Figure 9.5 can be carried out only once.
In the case of periodic disturbance, we can avoid prediction in the loop,
just like the SISO case. Use the same F (s) = G(s) as before but change H
to

H(s) = P̃ (s)e−τ̃h s ,

where τ̃h = k0 T and k0 is the smallest nonnegative integer that renders

τ̃h + Lij > 0 for all i and j. Such a H then has no pure prediction.
322 9. Dynamic Disturbance Decoupling

Example 9.2.4. Consider the well-known Wood/Berry binary distillation col-

umn plant (Wood and Berry, 1973):
 
12.8e−s −18.9e−3s
16.7s + 1 21.0s + 1 
P (s) =  .
6.60e−7s −19.4e−3s
10.9s + 1 14.4s + 1
The PI controller given by Luyben (1986b) is
 
1
0.375(1 + ) 0
 8.29s 
C(s) =  .
1 
0 −0.075(1 + )
23.6s
With model reduction, we obtain
 
0.41s4 + 0.65s3 + 0.41s2 + 0.04s + 0.001 e2s −0.27s4 − 0.43s3 − 0.27s2 − 0.02s − 0.001 e2s
 s4 + 0.33s3 + 1.17s2 + 0.18s + 0.008 s4 + 0.30s3 + 1.16s2 + 0.14s + 0.005 
P −1 ≈ P̃ = 

,

0.18s4 + 0.29s3 + 0.18s2 + 0.02s + 0.001 e−2s −0.23s4 − 0.37s3 − 0.23s2 − 0.02s − 0.001 e4s
4 3 2 4 3 2
s + 0.35s + 1.18s + 0.20s + 0.01 s + 0.32s + 1.17s + 0.16s + 0.006

where there are three pure predictors e2s , e2s and e4s in P̃ (s). According to
(9.70), we have
L = 4.

It then follows from (9.71) that P̃ (s) is rewritten as

P̃ (s) = diag {e4s , e4s }·
 
0.41s4 + 0.65s3 + 0.41s2 + 0.04s + 0.001 e−2s −0.27s4 − 0.43s3 − 0.27s2 − 0.02s − 0.001 e−2s
 s + 0.33s + 1.17s + 0.18s + 0.008
4 3 2 4 3 2
s + 0.30s + 1.16s + 0.14s + 0.005 
 .
 
0.18s4 + 0.29s3 + 0.18s2 + 0.02s + 0.001 e−6s −0.23s4 − 0.37s3 − 0.23s2 − 0.02s − 0.001
4 3 2 4 3 2
s + 0.35s + 1.18s + 0.20s + 0.01 s + 0.32s + 1.17s + 0.16s + 0.006

The Nyquist arrays of P −1 (s) and P̃ (s) are depicted in Figure 9.14, exhibiting
good matching of their frequency responses.
To approximate e4s , we set τv = L/2 = 2 and get Kvmax = 1.52. By setting
0.38
Kv = Kvmax /4 = 0.38, V (s) = 2s+1 is obtained. A unit step change is made
at t = 0 for the 1st set-point and at t = 200 for 2nd set-point, respectively, and
1
a step disturbance of size 1 is further introduced through Gd = 20s+1 to the
system at t = 400. The resulting output responses are shown in Figure 9.15.
It can be seen that with the add-on control to the conventional feedback
control, the system has achieved a better load disturbance rejection.
To evaluate the control scheme of Figure 9.7 for periodic disturbances,
the following periodic disturbance instead of a step disturbance is introduced
at t=400:
d(t) = sin(0.1t) + 0.3 sin(0.2t) + 0.1 sin(0.3t) + 0.1 sin(0.4t)
2π
with the period of T = 0.1 = 62.8 and G(s) = 1. τ̃h = k0 T = 62.8 is adopted
with k0 = 1. The system responses are shown in Figure 9.16. It is observed
 9.3 Notes and References 323

0.5 0

−0.05
0.4
−0.1
0.3 −0.15

0.2 −0.2

−0.25
0.1
−0.3

0 −0.35
−1 −0.5 0 0.5 −0.2 0 0.2 0.4 0.6

0.05 0.6

0
0.4

−0.05
0.2
−0.1

0
−0.15

−0.2 −0.2
0 0.1 0.2 0.3 0.4 −0.2 −0.1 0 0.1 0.2 0.3

−1
Fig. 9.14. Nyquist Array of P (s) and P̃ (s) for Example 9.2.4
(—— P −1 (s), · · · P̃ (s))

that our output response approaches zero asymptotically, as expected, while

the conventional feedback control has to sustain quite big steady state dis-
turbance response. ♦

In this section, a control scheme for disturbance rejection is presented for

time-delay processes. It can compensate for the disturbance while retaining
the input-output servo response of the conventional feedback system. Stabil-
ity and disturbance rejection performance are discussed.

9.3 Notes and References

The disturbance decoupling problem using a state feedback was considered in

a geometric setting by Basile and Marro (1969) and Wonham (1979). Equiv-
alent frequency domain solvability conditions were given by Bhattacharyya
(1980; 1982). In the context of output feedback, the problem was solved
by Akashi and Imai (1979) and Schumacher (1980). However, they did not
consider the issue of stability. The disturbance decoupling by measurement
feedback with stability or pole placement was solved for the first time by
Willems and Commault (1981). This problem was also considered from an
324 9. Dynamic Disturbance Decoupling

1.5

1
d
0.5

0
0 100 200 300 400 500 600

1.5

1
y1

0.5

0
0 100 200 300 400 500 600

1.5

1
y2
0.5

0
0 100 200 300 400 500 600

Fig. 9.15. Control Performance for Example 9.2.4

(—— with Disturbance Observer, · · · without Disturbance Observer)

input-output viewpoint by Kucera (1983) for single variable system and by

Pernebo (1981) for multivariable systems. Section 1 of this chapter is based
on the exiting results with some elaborations and modifications. These re-
sults on linear systems have recently been extended to nonlinear systems by
Andiarti and Moog (1996) and Xia (1996).
The disturbance observer was introduced by Ohnishi (1987). It was further
refined by Umeno and Hori (1991). But they considered delay-free systems
only. Section 2 of this chapter extends them to the delay case and is from
Zhang (2000).
 9.3 Notes and References 325

d 0

−1

−2
0 100 200 300 400 500 600 700 800

1.5

y1 1

0.5

0
0 100 200 300 400 500 600 700 800

2
y2

0
0 100 200 300 400 500 600 700 800

Fig. 9.16. Control performance for Example 9.2.4 under Periodic Disturbance
(—— with Disturbance Observer, · · · without Disturbance Observer)
10. Asymptotic Disturbance Decoupling

Dynamic disturbance decoupling, as discussed in the preceding chapter, aims

to decouple the disturbance from the output dynamic transient completely
and thus inevitably results in some restrictive conditions on its solvability. For
instance, the disturbance may not be measurable so that feedforward control
is not applicable; And the plant inverse may not be stable or realizable,
which may limit applicability or accuracy of the disturbance observer. A
natural next target after dynamic disturbance decoupling would be static
or asymptotic disturbance decoupling, that is, to decouple the disturbance
from the plant output at the steady-state only. For measurable disturbances,
asymptotic disturbance decoupling problem is straightforward. Rather, we
will focus on the unmeasurable disturbance case in this chapter. A general
case of disturbances will be considered in Section 1 whereas a specific yet
common case of periodic disturbances is treated in Section 2.

10.1 General Disturbances: Internal Model Principle

Asymptotic disturbance decoupling is also called asymptotic disturbance re-

jection or asymptotic regulation. A closely related problem is asymptotic
reference tracking since both the disturbance and reference are exogenous
signals to the feedback system. Thus, in this section, we will develop the
theory of asymptotic tracking and regulation for multivariable systems in a
general setting.
Let the plant be described by a transfer matrix model:

y(s) =Gu (s)u(s) + Gd (s)d(s), (10.1)

ym (s)=Gm (s)y(s), (10.2)

where y(s) is the m-dimensional output to be controlled, u(s) the l-dimen-

sional control input, d(s) the p-dimensional disturbance to be rejected, ym (s)
the m-dimensional measurement of the output y, to be used as the feed-
back signal for the compensator. Proper rational matrices Gu , Gd , and Gm
with appropriate dimensions represent plant, disturbance channel and sensor
models, respectively. We assume that the sensor Gm is stable, which is al-
most always the case in practice and therefore not restrictive. For closed-loop

Q.-G. Wang: Decoupling Control, LNCIS 285, pp. 327-356, 2003.

Springer-Verlag Berlin Heidelberg 2003
328 10. Asymptotic Disturbance Decoupling

stabilizability, the series connection of the sensor and plant, Gm · Gu , should

have no unstable pole-zero cancellations. Let the compensator be
u(s) = Gc (s)e(s), (10.3)
where Gc (s) is a proper rational matrix, e(s) = r(s) − ym (s) is the error,
and r(s) is the m-dimensional reference input to be tracked by the control
system. The overall system considered here is depicted in Figure 10.1.
Asymptotic regulation means limt→∞ y(t) = 0 in the case of r(s) = 0, for
any d(s) generated by
˜
d(s) = Ḡd (s)d, (10.4)
where d˜ is a real vector and Ḡd a proper rational matrix. On the other hand,
asymptotic tracking means limt→∞ e(t) = 0 in the case of d(s) = 0 for any
r(s) generated by
r(s) = G̃r (s)r̃, (10.5)
where r̃ is a real vector and G̃r a proper rational matrix. Stability of the
system is obviously a prerequisite for asymptotic regulation and/or tracking
to take place.

~ d (s )
d Gd (s ) Gd (s )

~ +
r ~ r (s ) e(s ) u(s ) y (s )
Gr ( s ) Gc (s ) Gu (s )
+ +
-
y m (s )
Gm (s )

Fig. 10.1. Servo Control

For a (real) polynomial matrix P (s) of full rank, we denote by σ(P(s))

the set of all c ∈ C, the complex plane, such that P (c) is of rank defect.
By the polynomial division and partial fraction expansion, a rational matrix
G(s) can be written uniquely as the sum of a polynomial matrix PG (s) plus
two strictly proper rational matrices G+ (s) and G− (s):
G(s) = PG (s) + G+ (s) + G− (s),
where all the poles of anti-stable G+ (s) and stable G− (s) lie, respectively, in
C+ , the closed right half of the complex plane (RHP), and C− , the open left
half of the complex plane (LHP). Apply this decomposition to G̃r and
G̃d := Gd Ḡd . (10.6)
 10.1 General Disturbances: Internal Model Principle 329

By the final value theorem of the Laplace transform, one can conclude that for
a stable system, the steady-state behavior of y(t) and e(t) is independent of
the stable parts, G̃− −
d and G̃r , of G̃d and G̃r . We can therefore assume without
loss of generality that any stable modes, which may be initially associated
with G̃d and G̃r , have been removed when defining G̃d and G̃r , in other
words, all the poles of both G̃d and G̃r can be assumed to lie in C+ .
We now formulate a general servocompensator problem with internal sta-
bility as follows. (SP): given a system in (10.1)-(10.3) and the disturbance
and reference in (10.4)-(10.5), as depicted in Figure 10.1 with Gu , Gm , G̃d ,
and G̃r all known and proper rational, Gu and Gm having no RHP pole-zero
cancellations and G̃d and G̃r being anti-stable (i.e., having the RHP poles
only), determine a proper compensator Gc such that the system of (10.1)-
(10.3) is internally stable and e(t) asymptotically tends to zero for any real
vectors d˜ and r̃ . In particular, the case of d˜ = 0 is called tracking problem
with internal stability (TP), and the case of r̃ = 0 and e replaced by y is
called regulation problem with internal stability (RP).

10.1.1 Preliminaries

Recall from Chapter 2 that two polynomial matrices, D ∈ Rm×m [s] and
N ∈ Rm×l [s], are left coprime (implicitly in C) if Rank[D(s) N (s)] =
m, f or all s ∈ C. This can be generalized as follows.
Definition 10.1.1. Two polynomial matrices, D ∈ Rm×m [s] and N ∈
Rm×l [s], are called left coprime in C+ , if

Rank[D(s) N (s)] = m, f or all s ∈ C+ .

Right coprimeness in C+ can be defined in a dual way.

Consider a cascade system G1 (s)G2 (s) with proper G1 (s) and G2 (s). They
can be expressed as coprime polynomial matrix fractions

G1 (s) = D1−1 (s)N1 (s), G2 (s) = N2 (s)D2−1 (s), (10.7)

respectively. For the sake of simplicity the indeterminate s may be dropped

as long as it causes no confusion. By Chapters 2 and 3, it is trivial to establish
the following lemma.
Lemma 10.1.1. For G1 and G2 in (10.7), the following are equivalent:
(i) G1 · G2 has no RHP pole-zero cancellations;
(ii) Neither D1−1 ·N1 N2 nor N1 N2 ·D2−1 has any RHP pole-zero cancellations;
(iii) Both fractions, D1−1 N1 N2 and N1 N2 D2−1 , are coprime in C+ ;
(iv) There are stable rational matrices A, B, X and Y such that
D1 A + N1 N2 B = I,
XN1 N2 + Y D2 = I;
330 10. Asymptotic Disturbance Decoupling

(v) The system is internally stabilizable, i.e. there is a proper compensator

which internally stabilizes G1 G2 .

Lemma 10.1.2. Let G = G1 G2 with left coprime fractions G = D−1 N ,

G1 = D1 N1 and G2 = D2−1 N2 , let R be a polynomial matrix and R−1 anti-
−1

stable. i.e. σ(R) ⊂ C+ and G2 be stable, i.e., σ(D2 ) ⊂ C− . Then if R is a

right divisor of D, it is also a right divisor of D1 .
Proof. We prove by contradiction. Suppose that D1 R−1 is not a polynomial
matrix and represented by a right coprime fraction D̄1 R̄−1 with a non-empty
σ(R̄) ⊂ C+ . Since D = P R for some polynomial matrix P , then we have
G = G1 G 2
(P R)−1 N = D1−1 N1 D2−1 N2
N1 D2−1 N2 = D1 R−1 P −1 N = D̄1 (P R̄)−1 N
Since P R and N are coprime, so are P R̄ and N . This together with coprime-
ness of R̄ and D̄1 implies that the RHP set σ(R̄) ⊂ σ(D2 ), which contradicts
the assumed stability of G2 . The proof is completed.

For a nonsingular polynomial matrix D, it can be factorized into D =
D− D+ such that σ(D−1 ) ⊂ C− and σ(D+ ) ⊂ C+ . Such a factorization
can be done, for example, by determining the Smith form of D and then
factorizing every invariant polynomial as the product of a stable polynomial
and an anti-stable polynomial. Let rational matrices G1 and G2 have the same
number of rows with coprime fractions G1 = D1−1 N1 and G2 = D2−1 N2 . The
above factorization is applied to D2 , i.e., D2 = D2− D2+ . G1 is said to have
the unstable model of G2 if D2+ is a right divisor of D1 , i.e., D1 (D2+ )−1 is a
polynomial matrix. Applying Theorem 9.1 of Chapter 9 to our case of Gf = 0
(comparing Figure 9.1 with Figure 10.1) yields the following.
Lemma 10.1.3. The system (10.1)-(10.3) is internally stable if and only if
the following hold.
(i) Gu has the unstable model of Gd ; and
(ii) Gc internally stabilizes Gm Gu , which is equivalent to that Gm · Gu
has no RHP pole-zero cancellations and Gc stabilizes Gm Gu .
The notion of skew primeness was introduced by Wolovich (1978). A full
row rank polynomial matrix N ∈ Rm×l [s] and a nonsingular polynomial
matrix D ∈ Rm×m [s] are said to be externally skew prime (implicitly in C)
if there are a full row rank N̄ ∈ Rm×l [s] and a nonsingular D̄ ∈ Rl×l [s] such
that DN = N̄ D̄, with D and N̄ left coprime (in C) and N and D̄ right
coprime (in C). We will need its generalization as follows.
Definition 10.1.2. Let a full row rank polynomial matrix N ∈ Rm×l [s] and
a nonsingular polynomial matrix D ∈ Rm×m [s] be given. N and D are said to
be externally skew prime in C+ if there are a full row rank polynomial matrix
N̄ ∈ Rm×l [s] and a nonsingular polynomial matrix D̄ ∈ Rl×l [s] such that
 10.1 General Disturbances: Internal Model Principle 331

DN = N̄ D̄, (10.8)

where D and N̄ are left coprime in C+ , and N and D̄ right coprime in C+ .

Such (N̄ , D̄) is refereed to as an internally skew prime pair of (N, D) in C+ .
In our proofs below, we will frequently make use of Lemma 5.2 of Chapter
5, and thus repeat here for ease of reference.
Lemma 10.1.4. Let A, B, C, D be all polynomial matrices such that
AB −1 C = D. If B −1 C (resp. AB −1 ) is coprime, then AB −1 (resp. B −1 C)
is a polynomial matrix.

10.1.2 Solvability

We will solve (RP) first, and subsequently apply its solution to (TP) and (SP).
Given Gu , Gm and G̃d , we can obtain polynomial fraction representations as
follows:

Gu =Du−1 Nu , (10.9)
−1
Gm =Dm Nm , (10.10)
−1
Du G̃d =Dd Nd , (10.11)
Nm Du−1 =D1−1 N1 , (10.12)

where all fractions on the right-hand sides are coprime and Dd−1 is anti-stable.
Theorem 10.1.1. (RP) is solvable if and only if
(i) Gu has the unstable model of Gd ;
(ii) N1 and Dd are right coprime in C+ ; and
(iii) Nu and Dd are externally skew-prime in C+ .
Proof. Sufficiency: The external skew-primeness of Nu and Dd in C+ means
that there are D̄d and N̄u such that Dd−1 N̄u = Nu D̄d−1 and both fractions
are coprime in C+ . Take

Gc = D̄d−1 Ḡc .

Write
Gm Gu D̄d−1 = [(D1 Dm )−1 N1 Nu ][D̄d−1 ].
Our problem formulation assumes that Gm · Gu has no RHP pole-zero can-
cellations. Internal stability will follow from Lemma 10.1.3 if Gm Gu D̄d−1 is
internally stabilizable, or equivalently, by Lemma 10.1.1, if (N1 Nu )D̄d−1 is
right coprime in C+ . Since both (N1 , Dd ) and (Nu , D̄d ) are right coprime in
C+ , there exist stable rational matrices X1 , Y1 , X2 , Y2 such that
X1 N1 + Y1 Dd = I,
X2 Nu + Y2 D̄d = I.
332 10. Asymptotic Disturbance Decoupling

Post-multiplying the first by Nu and substituting to the second with replace-

ment of Dd Nu = N̄u D̄d give
X2 (X1 N1 Nu + Y1 N̄u D̄d ) + Y2 D̄d = I,
or
(X2 X1 )N1 Nu + (Y2 + X2 Y1 N̄u )D̄d = I.
The last equation implies that N1 Nu and D̄d are right coprime in C+ . Let Ḡc
be such that it stabilizes Gm Gu D̄d−1 , which is always possible (Chapter 3).
Consider now the steady state error. The transfer matrix between y and
d˜ is obtained from Figure 10.1 as
H = (I + Gu Gc Gm )−1 G̃d .
Let Ḡc Gm = N2 D2−1 be a right coprime fraction. It follows that
H = (I + Du−1 Nu D̄d−1 N2 D2−1 )−1 G̃d
= D2 (Dd Du D2 + N̄u N2 )−1 Nd
:= D2 P −1 Nd .
Stabilization of Gm Gu D̄d−1 by Ḡc ensures σ(P ) ⊂ C− . Hence, H is stable
and by the final-value theorem of the Laplace transform, lim t→∞ y(t) = 0 for
any d.˜ (RP) is thus solvable.
Necessity: If (RP) is solvable, then the system is internally stable and H
is stable. Condition (i) is obviously necessary, in view of Lemma 10.1.3. Write
Nu Gc Gm = D3−1 N3 where D3 and N3 are left coprime. It follows that
H(s) = [D3 Du + N3 ]−1 D3 Dd−1 Nd .
Internal stability of the system implies that [D3 Du +N3 ]−1 is stable. Note that
Dd−1 is anti-stable and Dd and Nd are right coprime. As a result, stability of
H implies that D3 Dd−1 must be a polynomial matrix, or Dd is a right divisor
of D3 . It follows from Lemma 10.1.2 that for stable Gm as assumed in the
formulation, Dd is also a right divisor of D4 , where Nu Gc = D4−1 N4 is a left
coprime fraction.
Look now at Nu Gc . Let Gc = Dc−1 Nc be coprime and D̃c−1 Ñu be a coprime
fraction of Nu Dc−1 . Note that closed-loop stability implies that Nu Dc−1 has
no RHP pole-zero cancellations and
σ + (D̃c ) = σ + (Dc ), (10.13)
where σ + (D) denotes the set of all c in C+ such that D(c) is singular. By
our construction, (D̃c−1 )(Ñu Nc ) is also a coprime fraction of Nu Gc so that

(D̃c−1 )(Ñu Nc ) = D4−1 N4 ,

D̃c = U D4 = D̃c1 Dd (10.14)

 10.1 General Disturbances: Internal Model Principle 333

for a unimodular U and a polynomial D̃c1 . It then follows from Nu Dc−1 =

D̃c−1 Ñu that

D̃c1 Dd Nu = Ñu Dc . (10.15)

−1
D̃c1 Ñu must be coprime because otherwise Dd−1 D̃c1
−1
Ñu = D̃c−1 Ñu (coprime
−1
by construction) could not be coprime. Let N̄u D̄c1 be a dual right coprime
−1
fraction of left coprime D̃c1 Ñu so that

σ + (D̃c1 ) = σ + (D̄c1 ). (10.16)

Equation (10.15) can be rewritten as

Dd Nu = N̄u D̄d ,

where
−1
D̄d := D̄c1 Dc (10.17)

must be, by Lemma 10.1.4, a polynomial matrix, due to right coprimeness of

−1
N̄u D̄c1 . Then, Nu and Dd are externally skew-prime in C+ if both (Dd , N̄u )
and (Nu , D̄d ) are coprime in C+ . Dd and N̄u must be coprime (then coprime
in C+ as well) because otherwise in

Dd−1 N̄u D̄c1

−1
= Dd−1 D̃c1
−1
Ñu = (D̃c1 Dd )−1 Ñu = (D̃c )−1 Ñu ,

the last fraction which is coprime by construction could not be coprime. It

follows from (10.17), (10.16), (10.13) and (10.14) that

σ + (D̄d ) = σ + (Dc ) − σ + (D̄c1 )

= σ + (Dc ) − σ + (D̃c1 )
= σ + (D̃c ) − σ + (D̃c1 )
= σ + (Dd ).

This implies that Nu and D̄d are right coprime in C+ , because otherwise
Dd−1 N̄u being equal to Nu D̄d−1 could not be coprime.
One notices
−1
Gm Gu Gc = D m (Nm Du−1 )(Nu Dc−1 )Nc
−1 −1
= Dm D1 N1 D̃c−1 Ñu Nc
= Dm D1 (N1 Dd−1 )D̃c1
−1 −1 −1
Ñu Nc .

It follows from Lemma 10.1.1 that Dd and N1 must be right coprime in C+

in order to preserve internal stabilizability. The proof of Theorem 10.1.1 is
therefore completed.

A physical interpretation can now be given to Theorem 10.1.1. Let

G̃d = D̃d−1 Ñd be a coprime fraction. Alternatively, G̃d = Du−1 Du G̃d =
334 10. Asymptotic Disturbance Decoupling

Du−1 Dd−1 Nd = (Dd Du )−1 Nd , and the last representation is also a fraction
of G̃d . So there is a polynomial matrix L such that Dd Du = LD̃d . In the
light of Theorem 10.1, in order to solve (RP) we must set Gc = D̄d−1 Ḡc . One
sees
Gu Gc = Du−1 Nu D̄d−1 Ḡc
= Du−1 Dd−1 N̄u Ḡc
= (Dd Du )−1 N̄u Ḡc
= (D̃d−1 )(L−1 N̄u Ḡc ),
as shown in Figure 10.2. It is clear that when (RP) is solved, an exact du-
plication of the inverse of the denominator matrix of the exogenous system
must be present in the internal loop at the same summation junction as the
exogenous signals enter the loop. This is known as ”internal model principle”.

~ ~ ~
d N d ( s) Dd−1 ( s )

+
~ y (s )
−1
L N u Gc ( s ) Dd−1 ( s )
+

Gm (s )
Fig. 10.2. Internal Model Principle

To apply Theorem 10.1.1 to (TP). Figure 10.1 is redrawn into Figure 10.3.
Then, (TP) is obviously equivalent to (RP) with Gu , Gm and Gd replaced
by Gm Gu , I and I respectively. Let
−1
Gm Gu = Dmu Nmu ,
Dmu G̃r = Dr−1 Nr ,
where all fractions on the right hand sides are coprime. Noting that Gm Gu
and I cannot have any pole-zero cancellations and I has no unstable modes at
all so that they trivially meet condition (i) of Theorem 10.1, we obtain from
Theorem 10.1.1 the following corollary for solvability of asymptotic tracking
problem.

Corollary 10.1.1. (TP) is solvable if and only if Nmu and Dr are exter-
nally skew prime in C+ .
 10.1 General Disturbances: Internal Model Principle 335

~
r ~ r
Gr ( s )

+
e(s )
Gc (s ) Gu (s ) Gm (s )
-

Fig. 10.3. Tracking Problem

For the general (SP), we combine Theorem 10.1 and Corollary 10.1 as
follows.
Corollary 10.1.2. (SP) is solvable if and only if
(i) Gu has the unstable model of Gd ;
(ii) N1 and Dd are right coprime in C+ ;
(iii) Nu and Dd are externally skew-prime in C+ ; And
(iv) Nmu and Dr are externally skew-prime in C+ .

10.1.3 Algorithms
Based on Theorem 10.1.1, a compensator which solves (RP) can be con-
structed as follows. Given transfer matrices Gu , Gm and G̃d ,
(i) determine left coprime fractions in (10.9)–(10.12);
(ii) check the solvability conditions of Theorem 10.1. If they are met, con-
struct D̄d ; and
(iii) design a compensator Ḡc that stabilizes Gm Gu D̄d−1 and meets other
control requirements if any.

Coprime Fractions. We wish to factorize a m × l proper transfer matrix

G(s) in the form of
G = D−1 N, (10.18)
where D and N are left coprime polynomial matrices. Rewrite (10.18) as
 

 G
DN = 0.
−I
It can be seen that a left coprime fraction of G can be determined by finding a
minimal basis for the left null space of [GT −I]T . This is a standard minimal
design problem and its algorithm is available (Kung and Kailath,1980).

Right Coprimeness. One of the solvability conditions for (PR) is the

right coprimeness of some matrix pair. A generalized Sylvester resultant sk
336 10. Asymptotic Disturbance Decoupling

(Bitmead et al 1978) can be used to check the coprimeness, where sk of

two polynomial matrices D(s) and N (s) of order k consists of the coefficient
matrices of D(s) and N (s). The coprimeness of D(s) and N (s) is determined
by the rank information of sk .

Stabilizing Compensator. If D̄d has been constructed somehow (see

below for its construction), the feedback loop can be stabilized by an appro-
priate choice of the remaining unspecified part of compensator, namely Ḡc .
The actual selection of an appropriate stabilizing compensator is now a stan-
dard problem in the linear system theory, which has a variety of constructive
solutions. No further discussion is needed.

Skew-Primeness. The external skew-primeness of two polynomials ma-

trices N and D is crucial for the solvability of (RP), and it is also involved
in construction of D̄ for the compensator design. We now present a construc-
tive procedure for direct computation of D̄ from N and D for a case which
is encountered very often.
Let an m × m nonsingular polynomial matrix D and an m × l full row
rank polynomial matrix N be given. The case which we want to consider is
that σ(D) and σ(N ) are distinct, i.e., σ(D) ∩ σ(N ) = ∅. It is well known
(Chapter 2) that a polynomial matrix can be transformed to its Smith form
by elementary operations. In other words, there are unimodular Li and Ri , i =
1, 2, 3, such that
D = L 1 S1 R 1 , (10.19)
N = L 2 S2 R 2 , (10.20)
DN = L3 S3 R3 , (10.21)
where Si , i = 1, 2, 3 are in the Smith form:

S1 = Λ 1 , S2 = [Λ2 0], S3 = [Λ3 0]

for diagonal Λi . Recall that a determinantal divisor of order k of a polynomial

matrix is the monic greatest common divisor of all its minors of order k. Let
d1k , d2k and d3k be the determinantal divisors of order k of D, N and DN ,
respectively. Then Λi are given by

Λi = diag{λi1 , λi2 , · · · , λim }, i = 1, 2, 3,

where λik = dik d−1

i,k−1 , k = 1, 2, . . . , m, with di0 = 1 by convention.

Lemma 10.1.5. For nonsingular D and full row rank N with σ(D) ∩
σ(N ) = ∅, the Smith form of DN is the product of those of D and N , that
is, S3 = S1 S2 .

Proof. From (10.19) and (10.20), DN = L1 (S1 R1 L2 S2 )R2 , the Smith form of
DN is the same as that of S1 R1 L2 S2 because L1 and R2 are unimodular. It
is therefore sufficient to prove that the Smith form of S1 M S2 is S1 S2 , where
 10.1 General Disturbances: Internal Model Principle 337

M is unimodular. Let Q = S1 M S2 . Note that the determinantal divisors of

order k for S1 , S2 and Q are, respectively, the same as those for D, N and
DN , and they are still denoted by d1k , d2k and d3k ,respectively.
Set Q = S1 Q1 for Q1 = M S2 . By the Binet-Cauchy formula, we can
express a minor of Q of order k as a linear combination of minors of S1 of
the same order. It follows that
d3k d−1
1k = p1k , k = 1, 2, . . . , m, (10.22)
for some polynomial p1k . Similarly, by setting Q = Q2 S2 where Q2 = S1 M ,
one gets
d3k d−1
2k = p2k , k = 1, 2, . . . , m, (10.23)
for some polynomial p2k . For a fixed k, partition M into [Mk Mm−k ], where
Mk is the first k columns of M . It can be seen from Q = S1 M S2 that gk mk d2k
is a minor of order k of Q, where mk is any minor of order k of Mk and gk is
the product of the diagonal elements of a submatrix of S1 corresponding to
mk and is a polynomial. The greatest common divisor of all mk must be a
non-zero real constant, because otherwise a Laplace expansion of the det(M )
would show that M were not unimodular. Therefore the greatest common
divisor of all gk mk d2k can be expressed as f1k d2k where f1k is a polynomial
and σ(f1k ) ⊂ σ(S1 ) = σ(D). By definition of d3k , we have
f1k d2k d−1
3k = p̄2k , k = 1, 2, . . . , m, (10.24)
where p̄2k is a polynomial. Similarly, one can get
f2k d1k d−1
3k = p̄1k , k = 1, 2, . . . , m, (10.25)
where f2k and p̄1k are polynomials and σ(f2k ) ⊂ σ(S2 ) = σ(N ). Factorize
d3k into d3k = dD N D N
3k d3k where σ(d3k ) ⊂ σ(D) and σ(d3k ) ⊂ σ(N ). Then we
see that (10.22) and (10.25) are satisfied simultaneously if and only if both
−1
dD D −1
3k d1k and d1k (d3k ) are polynomial. This is possible only if dD 3k = d1k .
N
Similarly, we have d3k = d2k . In view of these relations, we obtain
λ3k = d3k (d3,k−1 )−1 = dD D
3k (d3,k−1 )
−1 N
d3k (dN
3,k−1 )
−1

= (d1k d−1 −1
1,k−1 )(d2k d2,k−1 )
= λ1k · λ2k ,
which implies that the Smith form of Q = S1 M S2 is S1 S2 , The proof is thus
complete.

Theorem 10.1.2. Let a full row rank N ∈ Rm×l [s] and a nonsingular
m×m
D∈R [s] be given with σ(D) ∩ σ(N) = ∅. Then N and D are externally
skew prime and thus externally skew prime in C+ as well. Furthermore, D̄
and N̄ , defined by
 
S1 0
D̄ = R3 , (10.26)
0 I
N̄ = L3 S2 , (10.27)
338 10. Asymptotic Disturbance Decoupling

with I being an identity matrix of order (l − m), constitute an internally skew

prime pair of (N, D) in the normal sense as well as in C+ .
Proof. From Lemma 10.1.5, one can write

DN = L3 S3 R3 = L3 S1 S2 R3
= L3 [Λ1 Λ2 0]R3 
Λ1 0
= L3 [Λ2 0] R3
 0 I
S1 0
= L 3 S2 R3
0 I
:= N̄ D̄,

where N̄ = L3 S2 and D̄ = Blockdiag{S1 , I}R3 are both polynomial matrices.

Furthermore, D and N̄ must be a left coprime because otherwise there is a
non-unimodular common divisor P such that σ(P ) ⊂ σ(D) and σ(P ) ⊂
σ(N̄ ) = σ(S2 ) = σ(N ), which contradicts the fact that σ(D) ∩ σ(N ) = ∅.
Similarly, N and D̄ must be right coprime. Hence, N and D are externally
skew prime. Note the coprimeness in the normal sense, i.e., in C, of course,
implies coprimeness in C+ , a subset of C. N and D are thus externally skew
prime in C+ as well.

Example 10.1.1. For illustration of Theorems 10.2 and 10.1, consider the
headbox model of a paper machine as established in Wang (1984):
y(s) = Gu (s)u(s) + Gd (s)d(s),
where y = [y1 y2 ]T , y1 and y2 are the liquid level and pulp concentration
in the headbox, respectively; u = [u1 u2 ]T , u1 and u2 are the flow rates of
thick pulp and white water, respectively; d = [d1 d2 ]T , d1 and d2 are the
pulp concentrations of thick pulp and white water respectively; and
 1.5 1.5

2.6s+1 2.6s+1
Gu = 5.1 −1.4
,
2s+1 2s+1
 
0 0
Gd = 0.2 0.8
.
2s+1 2s+1

The sensor model is

 
1 0
Gm = 1
.
0 2s+1
Suppose that both disturbances to be rejected are step signals:
   
d˜1
s 1 d˜1
d(s) =   = ˜
:= Ḡd d.
d˜1 s d˜2
s

We construct
 10.1 General Disturbances: Internal Model Principle 339

 +  
0 0 0 0
+
G̃d: = (Gd Ḡd ) = 0.2 0.8
= 0.2 0.8
.
s(2s+1) s(2s+1) s s

Appropriate coprime fractions are obtained as

 −1  
2.6s + 1 0 1.5 1.5
Gu = := Du−1 Nu ,
0 2s + 1 5.1 −1.4
 −1  
1 0 10 −1
Gm = := Dm Nm ,
0 s+1 01
 −1  
10 0 0
Du G̃d = := Dd−1 Nd ,
0s 0.2(2s + 1) 0.8(2s + 1)
 −1  
2.6s + 1 0 10
−1
Nm D u = := D1−1 N1 .
0 2s + 1 01
Since σ(Nu ) = ∅, then σ(Dd ) ∩ σ(Nu ) = ∅. In view of Theorem 10.1.2, Dd
and Nu are externally skew-prime. It follows that
 
10
Dd = := S1 ,
0s
which is already in the Smith form, the Smith form for N11 is S2 = I and
  
10 1.5 1.5
D d Nu =
0s 5.1 −1.4
   
10 10 1.5 1.5
= := L3 S3 R3 .
01 0s 5.1 −1.4
By Theorem 10.1.2, we have N̄u = L3 S2 = I and
 
1.5 1.5
D̄d = S1 R3 = .
5.1s −1.4s

Obviously, Gm · Gu has no RHP pole-zero cancellations. Gd is stable so

that condition (i) of Theorem 10.1 is trivially met. It is readily seen that
Dd and N1 are right coprime. Hence, (RP) for this example is solvable. Any
stabilizer for Gm Gu D̄d−1 will solve (RP). ♦

In this section, a servocompensator problem with internal stability has

been resolved in a polynomial matrix setting for general multivariable sys-
tems. It has been shown that the primary condition of solvability is that
the plant numerator matrix and the denominator matrix of the product of
the plant denominator matrix and exogenous system transfer matrix are ex-
ternally skew-prime in the RHP. This important observation naturally leads
to an internal model principle in an explicit way. We have also presented a
procedure for designing a compensator and discussed relevant computational
algorithms.
340 10. Asymptotic Disturbance Decoupling

10.2 Periodic Disturbances: Virtual Feedforward

Control
The internal model principle (IMP) presented in the preceding section is an
elegant theory and has wide applications in industry. In particular, it is easy
to apply if the disturbance is of simple type such as step, ramp and sinusoidal
where the unstable modes of the disturbance Laplace transform are clear
and simple to users, and stabilization of the generalized plant consisting of
the original plant and the internal model is easy to achieve. However, there
do exist other cases for which the IMP is difficult or even impossible to
apply. For instance, some disturbances may be irregular. Others may not
have analytical expressions for its Laplace transforms, so one is unable to
extract the unable modes of the disturbance under consideration. Even when
such unstable modes can be obtained, it may be impossible, too difficult or
expensive to stabilize the resulting generalized plant. A class of disturbances
which falls into this last case is general periodic signals other than sinusoidal.
In practice, we often encounter the situation where the reference com-
mands to be tracked and/or disturbance to be rejected are periodic signals,
e.g., repetitive commands, operations for mechanical systems such as indus-
trial robots, or disturbances depending on the frequency of the power supply.
Disturbances acting on the track-following servo system of an optical disk
drive inherently contain significant periodic components that cause tracking
errors of a periodic nature. For such disturbances, the controller designed for
step type reference tracking and/or disturbance rejection will inevitably give
an uncompensated error.
Theoretically, a signal of period T will contribute an unstable factor,
(1 − eLs ), into the denominator of its Laplace transform. By the IMP, the re-
ciprocal of this factor which has an infinite unstable modes has to be inserted
into the feedback loop. But how to stabilize it remains an open problem in the
literature. This motivates us to look for some alternative solution to the IMP.
In this section, a new control scheme, called the virtual feedforward control,
is presented for asymptotic rejection of periodic disturbance. The periodic
disturbance is estimated when a periodic steady state error is detected, and
the virtual feedforward control is then activated to compensate for such a
disturbance.

10.2.1 The Scheme

Consider first the case of SISO linear time-invariant (LTI) continuous pro-
cesses. Let G(s) be the plant and K(s) the feedback controller. The proposed
scheme, called the virtual feedforward control scheme (VFC scheme for short),
is depicted in Figure 10.4, where d is supposed to be an unknown periodic
disturbance. Without the VFC controller, the proposed structure reduces to
an ordinary feedback control system and in face of non-zero periodic d, the
system will have non-zero steady state error. The intention of the proposed
method is to activate the VFC controller timely and to give an extra control
 10.2 Periodic Disturbances: Virtual Feedforward Control 341

signal v to compensate for the disturbance. Without loss of generality, the

reference r is assumed to be zero when the disturbance response in output is
addressed.

Fig. 10.4. Feedback System with VFC

It is well known that the steady state of the output y(t) of a stable system
G in response to a periodic input x(t) is periodic with the same period as
that of the input. With the standard convolution operation ∗, the response
of a stable system Y = G2 G1 X can be written as y(t) = g2 (t) ∗ g1 (t) ∗ x(t),
where g1 (t) and g2 (t) are impulse responses of G1 and G2 , respectively. For
simplicity, denote g2 (t) ∗ g1 (t) ∗ x(t) by (g2 ∗ g1 ∗ x)(t). For a signal f (t) with
a periodic steady state, let its steady state be f s (t), and the corresponding
Laplace transform be F s (s) = L{f s (t)}. Lemma 9.3 in Chapter 9 is needed in
our development here and repeated as follows (readers are referred to Chapter
9 for its proof).
Lemma 10.2.1. Under a periodic input x(t), the output steady state of the
system Y = G2 G1 X with stable G2 and G1 satisfies

(g2 ∗ g1 ∗ x)s (t) = [g2 ∗ (g1 ∗ x)s ]s (t),

equivalently, in the s-domain, there holds

(G2 G1 X)s = [G2 (G1 X)s ]s .

With this lemma in hand, we are now ready to develop the new method
for periodic disturbance rejection. Suppose that the process is in a normal
and stable feedback control (v = 0 in Figure 10.4) and the output has been in
a constant steady state before the disturbance acts on the system. Without
loss of generality, such a steady state is assumed to be zero. When a periodic
disturbance d comes to the system at t = 0, the process output y will become
342 10. Asymptotic Disturbance Decoupling

periodic with the same period after some transient. The periodic output can
be detected and its period be measured with high accuracy by monitoring
the process output.
At t = 0, the VFC controller has not been activated yet (v = 0 in Fig-
ure 10.4), and it follows from Figure 10.4 that Yd = D + GUd , or

D = Yd − GUd , (10.28)

where the subscription d indicates the signals’ response to d only. Note that
yd (t) and ud (t) as well as their steady states yds (t) and usd (t) are available. In
the time domain, (10.28) becomes

d(t) = yd (t) − (g ∗ ud )(t). (10.29)

Since the actual process transfer function G is not available, its model Ĝ has
to be used to estimate d(t), i.e.,
ˆ = yd (t) − (ĝ ∗ ud )(t),
d(t) (10.30)

or in the s-domain,

D̂ = Yd − ĜUd . (10.31)

Note that d(t) and d(t)ˆ have the periodic steady state ds (t) and dˆs (t) respec-
tively, since yd (t) and ud (t) are so and the system is stable. Now suppose that
the VFC controller v(t) is activated at t = Tv . Since the initial conditions of
the system at t = Tv are non-zero, we have to apply the Laplace transform
at t = 0 to get
1 G
Yv = D− V, (10.32)
1 + GK 1 + GK
where the subscription v indicates the signal’s response to both d and v. Let
1(t) be the unit step function. We set the virtual feedforward control signal
as

V (s) = [Ĝ−1 (s)D̂(s)]s e−sTv . (10.33)

Let
1
Qvd = Ĝ−1 (s)D̂(s), Hyd = ,
1 + GK
Equation (10.33) can be written in the time domain as

v(t) = [qvd (t − Tv )]s 1(t − Tv ). (10.34)

Substituting (10.33) into (10.32) gives

Yv = Hyd D − Hyd G(Ĝ−1 D̂)s e−sTv .

 10.2 Periodic Disturbances: Virtual Feedforward Control 343

It follows from Lemma 10.2.1, (10.28) and (10.31) that

Yvs
= [Hyd D − Hyd G(Ĝ−1 D̂)s e−sTv ]s
= [Hyd Ds ]s − [Hyd GĜ−1 D̂s e−sTv ]s
= [Hyd Yds − Hyd (GUd )s ]s − [Hyd GĜ−1 Yds e−sTv − Hyd GĜ−1 (ĜUd )s e−sTv ]s
= [Hyd (1 − GĜ−1 e−sTv )Yds ]s − [Hyd G(1 − e−sTv )Uds ]s . (10.35)

Choose

Tv = kT, (10.36)

where T is the period of d(t), yds (t) and usd (t), and k is an integer, so that
there hold

yds (t − Tv ) = yds (t), usd (t − Tv ) = usd (t). (10.37)

Thus, (10.35) reduces to

yvs (t) = {hyd (t) ∗ [1 − g(t) ∗ ĝ −1 (t)] ∗ yds (t)}s . (10.38)

which becomes

yvs (t) = 0, if g(t) = ĝ(t),

i.e., in the case of perfect model-process match.

Now, if there is a model mismatch, i.e., g(t) = ĝ(t), define

1 − GĜ−1
F  ,
1 + GK
and f as its impulse response. (10.38) is written as

yvs (t) = {f (t) ∗ yds (t)}s . (10.39)

Introduce the following norm for a periodic signal x(t) with the period of T :

T
x(t) = x2 dt,
0

and express yds (t) by its Fourier series:

∞
yds (t) = yd0 + Ci cos(ωi t + φi )
i=1

with ωi = 2πi/T . We have

344 10. Asymptotic Disturbance Decoupling

T ∞
yds (t)2 = [yd0 + Ci cos(ωi t + φi )]2 dt. (10.40)
0 i=1

Noting the orthogonality of cos(ωi t) and cos(ωi t+φi ) on the interval t ∈ [0, T ],
(10.40) gives
∞
T
yds (t)2 = T yd0
2
+ Ci2 . (10.41)
2 i=1

Similarly, we have
T ∞
yvs (t)2 = [F (0)yd0 + Ci |F (jωi )| cos(ωi tφi + ∠F (jωi ))]2 dt
0 i=1
∞
T
= F 2 (0)T yd0
2
+ Ci2 |F (jωi )|2 . (10.42)
2 i=1

If
|F (jωi )| < 1, i = 1, 2, · · · , (10.43)
it follows from (10.42) that
yvs (t) < yds (t). (10.44)
In the VFC scheme, the extra control v is activated according to (10.34),
(10.30) and (10.36). This v is independent of the closed-loop system once
activated and acts as a “feedforward” control with the estimated disturbance
ˆ Therefore, stability of the closed-loop system can be guaranteed with the
d.
proposed VFC strategy, provided that the stability is taken into consideration
at the stage of controller design. In view of the above development, we obtain
the following theorem:
Theorem 10.2.1. A stable feedback system remains stable when the virtual
feedforward control signal v described by (10.34), (10.30) and (10.36) is in-
troduced, and the resultant output steady state in response to a periodic dis-
turbance satisfies

yvs (t) = 0, if Ĝ = G,
−1
yvs (t) < yds (t), if | 1−G(jω i )Ĝ(jωi )
1+G(jωi )K(jωi ) | < 1, i = 1, 2, · · · .

10.2.2 Implementation and Simulation

Let us discuss several practical issues in implementation of the VFC scheme.

Algorithm Execution. To implement the proposed VFC scheme shown

in Figure 10.4, the process output is monitored all the time. When a periodic
 10.2 Periodic Disturbances: Virtual Feedforward Control 345

behavior is observed, the VFC controller is activated: the steady state of the
disturbance d is estimated from (10.30) and v is computed from (10.34) and
applied to the system at Tv according to (10.36).

Non-minimum-phase system. In (10.34), the impulse response of the

model inverse is used to get the required dˆ and thus v for disturbance compen-
sation. For non-minimum-phase system, Ĝ−1 will give a divergent response.
However, note that only its steady state of dˆ is employed in the proposed
method to compute v. It is possible to extract this steady state part from
such a divergent response. For simplicity, consider a non-minimum-phase pro-
cess Y = GX with a single right-half-plant (RHP) zero at s = a > 0. Using
the partial fraction expansion, we have
β
X(s) = G−1 (s)Y (s) = + X t + X s, (10.45)
a−s
where X t denotes the stable part. The Laplace transform of the periodic
signal y(t) is given (Kuhfitting, 1978) by
T
1
Y (s) = y(t)e−st dt,
1 − e−sT 0
where T is the period. Applying the residue theorem yields
 
  1 T
−1 −at
β = lim (a − s)G (s) · y(t)e dt .
s→a 1 − e−aT 0

It can be seen from (10.45) that once the transient xt (t) dies out, the steady
state part xs (t) can be found by
xs (t) = [g −1 (t) ∗ y(t) − βeat ]s . (10.46)
The extension of the method to non-minimum-phase systems with multiple
RHP zeros is straightforward. In implementation, some numerical integration
technique (Sinha and Rao, 1991) has to be used to compute β, and this
inevitably gives a numerical error, which will accumulate and make η(t) 
g −1 (t) ∗ y(t) − βeat diverge. Thus, we only take one early period of η(t), when
η(t) becomes approximately periodic, and duplicate it by η s (t) = η s (t + T )
to generate xs (t) = η s (t) for all t  0.
For example, consider a non-minimum-phase process G(s) = −4s+1 5s+1 , and
1
suppose that a sinusoidal signal x(t) passes through 9s2 +4s+1 and then acts
on the process. With the suggested method, the steady state periodic signal
x̂s (t) is constructed and shown in Figure 10.5, where the solid line is x̂s (t)
and the dashed line is the actual x(t). It is clear that x̂s (t) is almost identical
to the steady state part of x(t).

Variation of v. The VFC signal v in (10.34) is equivalent to

−1
v(t) = [ĝ (t − Tv ) ∗ dˆs (t − Tv )]s 1(t − Tv ), (10.47)
346 10. Asymptotic Disturbance Decoupling

0.25

0.2

0.15

0.1

0.05

−0.05

−0.1

−0.15
0 5 10 15 20 25 30 35 40 45 50

−4s+1
Fig. 10.5. Non-minimum-phase System G(s) = 5s+1
(—— constructed x̂s (t), − − − actual x(t))

which is the periodic steady state of

v(t) = [ĝ −1 (t − Tv ) ∗ dˆs (t − Tv )]1(t − Tv ). (10.48)

Thus, (10.47) and (10.48) may be called static and dynamic VFC, respec-
tively. One may guess that the dynamic control (10.48) can improve distur-
bance rejection performance over the static one (10.47). Dynamic VFC and
static one have the same steady state, which causes the same steady state
output response. The difference between the signals generated by these two
schemes is its transient which depends on the plant and disturbance which
vary from the case to the case, and is hard to predict. This difference will
cause the transient difference in the output response and the latter is even
harder to predict. Thus, there is no guarantee that the dynamic VFC will
give an improved (dynamic) performance over the static VFC. What we see
from extensive simulation is that for some examples, dynamic VFC does
give a little improvement over static VFC, while for most examples, almost
no difference can be observed. Furthermore, the dynamic VFC is not appli-
cable to non-minimum-phase models. Therefore, we recommend (10.34) for
practical implementation for all stable models of both minimum-phase and
non-minimum-phase.

Adaptation. The proposed method can be extended to adaptively com-

pensate for disturbances with variable periods. The process output is mon-
itored all the time. When the periodic output is observed, v is applied and
 10.2 Periodic Disturbances: Virtual Feedforward Control 347

a new steady state with a suppressed error is expected after some transient.
Thereafter, one waits for a different output waveform to occur, and then the
VFC controller is de-activated (or removed) from the system. The procedure
is then repeated.
We now present some different scenarios to demonstrate the effectiveness
of the VFC scheme.
Example 10.2.1. Consider a process:
e−4s
G(s) =
(10s + 1)(2s + 1)
under a PID controller (Zhuang and Atherton, 1993):
0.2631
K(s) = 2.0260 + + 4.9727s. (10.49)
s
The external disturbance signal is introduced at t = 0:

d(t) = sin(0.1t) + 0.5sin(0.2t) + 0.3sin(0.3t) + 0.1sin(0.4t), (10.50)

2π
which has the period of T = 0.1 ≈ 62.83. The process output y is monitored
until the periodic response is detected at t = 150.0. The disturbance is esti-
mated via (10.30) and v(t) is computed from (10.34). v(t) is then applied at
Tv = 3T ≈ 188.5 to meet (10.36).
In the case of no model mismatch, i.e., Ĝ = G, the resultant responses are
shown in Figure 10.6. In the first plot, the disturbance is exhibited, where
the dashed line is d and the solid line is d. ˆ From this plot, it is clear that
the stationary part dˆs is identical to ds . In the second plot, the process
output is shown, where the solid line is from the proposed VFC scheme,
while the dashed line is from the normal feedback control. The effectiveness
of the VFC scheme is clear. Further, as stated in Theorem 10.3, the output
from the proposed control approaches zero asymptotically. Such a property
is guaranteed only when v is activated at Tv meeting (10.36). Suppose that v
is applied with an extra delay of t = 1, and there results a steady state error
as shown by the dotted line in the second plot of Figure 10.6.
The VFC scheme is independent of the feedback controller design, and
the output steady state approaches to zero provided that the condition of
Theorem 10.3 is satisfied. However, the output transient behavior inevitably
relies on the controller parameters. To see the controller effects, different
controller settings are used. Figure 10.7 plots the output responses, where
the solid line is from the controller (10.49), while the dashed line is from the
controller with the derivative part de-activated, i.e.,
0.2631
K(s) = 2.0260 + .
s
As expected, such a controller gives a slower response compared with that in
(10.49), though both produce a zero output asymptotically.
348 10. Asymptotic Disturbance Decoupling

2
1
d

0
−1

0 50 100 150 200 250 300 350 400

2
1
y

0
−1
0 50 100 150 200 250 300 350 400
5

0
u

−5

0 50 100 150 200 250 300 350 400

4
2
v

0
−2

0 50 100 150 200 250 300 350 400

Fig. 10.6. VFC for Example 10.2.1 without Model Mismatch

In practice, the process is not fully known and it has to be described by

a model. Suppose that the process is identified (Bi et al., 1999) as

1.03e−5.47s
Ĝ(s) = .
11.41s + 1

From the recorded y(t), v(t) is computed with Ĝ(s) and applied with Tv =
3T ≈ 188.5. The performance is shown in Figure 10.8 over t ∈ [0, 377.0],
where the solid line is with model mismatch, while the dashed line is for the
case without model mismatch.
 10.2 Periodic Disturbances: Virtual Feedforward Control 349

0
y

−1

−2

−3

0 100 200 300 400 500 600

Fig. 10.7. VFC for Example 10.2.1 with Different K(s)

(—— K(s) = 2.0260 + 0.2631
s
+ 4.9727s, − − − K(s) = 2.0260 + 0.2631
s
)

1.5
1
0.5
0
d

−0.5
−1
−1.5

0 100 200 300 400 500 600 700 800

0
y

−1

−2

−3
0 100 200 300 400 500 600 700 800

Fig. 10.8. VFC for Example 10.2.1 with Model Mismatch

350 10. Asymptotic Disturbance Decoupling

Suppose that at t = 6T ≈ 377.0, the disturbance is changed to D(s) =

1
15s+1 W (s), where

−1, if − T < t < 0,
w(t) =
1, if 0 < t < T.

Such a disturbance will change the waveform of the monitored y(t), and
a steady state periodic y(t) with enlarged error is observed. The adaptive
scheme is thus activated to de-activate v(t) at t = 7T ≈ 439.8. The virtual
feedforward block (with model mismatch) is activated again at t = 10T ≈
628.3 to compensate for the new disturbance, as shown in the solid line in
Figure 10.8. The response without the adaptive scheme, i.e., with the same
v(t) for the different d(t), is shown by the dashed line over t ∈ [377.0, 800.0]
in Figure 10.8.
In practice, measurement noise is generally present, and its level may be
measured by the noise-to-signal ratio (Haykin, 1989):

mean(abs(noise))
N SR = .
mean(abs(signal))

To demonstrate the robustness of the proposed method, the system is tested

again under a noise level with N SR ≈ 20% for the above scenario, and the
responses are shown in Figure 10.9. It is confirmed that the proposed method
is robust to measurement noise. ♦

Example 10.2.2. The proposed method can be applied to servo control prob-
lems for better disturbance rejection over the existing methods, say the repet-
itive control scheme. For demonstration, consider a servo system in Moon et
al. (1998) with the uncertain plant:

3.2 × 107 β0
G(s) = ,
s2 + α1 s + α0
where α1 ∈ [20, 27], α0 ∈ [45000, 69000], β0 ∈ [60, 100]; and the controller:

0.26(s + 1900)(s + 4100)(s + 11300)2

K(s) = .
(s + 100)(s + 19000)(s + 25100)(s + 31000)

We compare our method with the repetitive control scheme discussed in Moon
et al. (1998). Suppose that the system is in the steady-state before a distur-
bance is introduced at t = 0, the disturbance signal comprises a constant
signal, eight out-of-phase sinusoidal signals from the first to the eighth har-
monics of the spindle frequency, 30 Hz (Moon et al., 1998).
The coefficients of the model are supposed to be center values of each
interval, i.e.,

α1 = 23.5, α0 = 57000, β0 = 80. (10.51)

 10.2 Periodic Disturbances: Virtual Feedforward Control 351

1.5

0.5

0
d

−0.5

−1

−1.5

0 100 200 300 400 500 600 700 800

(a) disturbance

0
y

−1

−2

−3

0 100 200 300 400 500 600 700 800

(b) t = 0 ∼ 377.0: without model mismatch; t = 377.0 ∼ 800.0: with model

mismatch but without adaptation

0
y

−1

−2

−3

0 100 200 300 400 500 600 700 800

(c) with model mismatch and adaptation

Fig. 10.9. VFC for Example 10.2.1 with Noise N SR = 20%

For the case without model mismatch, i.e., the actual process is described
exactly by the model, the responses are shown in Figure 10.10, where the
solid line is from the proposed method, and the dashed line is from the
repetitive control scheme. Since the control signals from these two methods
are very close to each other, their difference ∆u is displayed in Figure 10.10,
352 10. Asymptotic Disturbance Decoupling

50
d (um)

−50
0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

0.02

0.01
y (um)

−0.01

−0.02
0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

0.05

0
u

−0.05
0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

−4
x 10
3

1
∆u

−1

−2

0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

Fig. 10.10. Comparison of VFC and Repetitive Control (—— VFC, − − − Repet-
itive)

where only the steady state parts are shown for illustration. The effectiveness
of the proposed method is obvious. Now we vary the plant coefficients within
the possible intervals:

α1 = 25.5436, α0 = 59770, β0 = 77.7881.

The responses are shown in Figure 10.11. The robustness of the proposed
method to the model uncertainty is noticed. ♦
 10.2 Periodic Disturbances: Virtual Feedforward Control 353

50
d (um)

−50
0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

0.04

0.02
y (um)

−0.02

−0.04
0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

0.05

0
u

−0.05
0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

−4
x 10
6

2
∆u

−2

−4

0.5 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

Fig. 10.11. VFC and Repetitive Control under Model Mismatch

(—— VFC, − − − repetitive)

10.2.3 Extension to the MIMO Case

We have so far considered the SISO case. The extension to the MIMO case
is straightforward. Suppose that the process P (s) is invertible and has been
stabilized by the controller K in a conventional feedback system. The same
scheme as in Figure 10.4 is exploited. The same derivation and implementa-
tion techniques as in the SISO case before applies with obvious modifications
to suit the square matrix case. The only thing to take note is that the inverse
354 10. Asymptotic Disturbance Decoupling

of the process G is needed in (10.34) to compute the VFC vector signal v(t)
and is usually highly complicated and not realizable. We can apply the model
reduction (Chapter 2) to G−1 (jω) to get its approximation G̃(s) in form of
g̃ij (s) = g̃ij0 (s)e−τij s . (10.52)
−τij s
where g̃ij0 (s) are proper rational functions. Some terms e may happen to
be a pure predictor but by choosing Tv = kT with a suitable k (see (10.36)),
we can always make all τij negative, leading to time delays terms e−τij s which
are realizable.
Example 10.2.3. Consider the process:
 12.8 −s −18.9 −3s 
16.7s+1 e 21.0s+1 e
G= 6.6 −7s −19.4 −3s
,
10.9s+1 e 14.4s+1 e

which is controlled by
 
0.949 + 0.1173/s + 0.4289s 0
K= .
0 −0.131 − 0.0145/s − 0.1978s
With model reduction, we obtain

G−1 ≈ G̃(s)
 
0.31s +0.62s3 +0.42s2 +0.05s+0.002 2s
4
0.089s4 +0.04s3 −0.19s2 −0.03s−0.0015 4s
s4 +0.25s3 +1.16s2 +0.21s+0.013 e s4 +0.36s3 +1.03s2 +0.17s+0.01 e
= .
0.14s4 +0.278s3 +0.18s2 +0.02s+0.0007 −2s 0.08s4 +0.03s3 −0.16s2 −0.03s−0.001 6s
s4 +0.26s3 +1.16s2 +0.22s+0.01 e s4 +0.38s3 +1.03s2 +0.19s+0.01 e

The following disturbance:

d(t) = sin(0.1t) + 0.5sin(0.2t) + 0.3sin(0.3t) + 0.1sin(0.4t)
is added to both outputs of the system at t = 0. The VFC controller is
activated at t = 3T (T is the period of the disturbance). The system time
responses with and without VFC are shown in Figure 10.12 with solid line
and dash line, respectively. Performance enhancement with VFC is obvious.
Its non-zero steady state error is caused by the approximation of the process
inverse. ♦
In this section, a control scheme, called the virtual feedforward control
(VFC), is presented for asymptotic disturbance rejection. The VFC is an
add-on function on top of normal feedback control, and it is activated when
a periodic disturbance is found to be present in the system, based on output
measurements and detection of output periodic behavior. It is shown that the
VFC can reject the periodic disturbances asymptotically under the perfect
plant modelling. Furthermore, the closed-loop stability is not affected by
the VFC and thus there is no design trade-off between disturbance rejection
and stability. The robustness of the control scheme to model mismatch is
analyzed. The effectiveness of the VFC method is sustained by simulation
and comparison.
 10.3 Notes and References 355

0.5

0
d

−0.5

−1

50 100 150 200 250 300 350 400 450 500

0.2

0.15

0.1
1
v

0.05

−0.05

50 100 150 200 250 300 350 400 450 500

0.1

0.05

−0.05
2
v

−0.1

−0.15

−0.2

50 100 150 200 250 300 350 400 450 500

1.4

1.2

0.8
y1

0.6

0.4

0.2

0
50 100 150 200 250 300 350 400 450 500

1.5
y2

0.5

0
50 100 150 200 250 300 350 400 450 500

Fig. 10.12. Disturbance Response for Example 10.2.3

(—— With VFC − − −Without VFC)

10.3 Notes and References

A basic requirement of a closed -loop control system is its ability to regu-

late against disturbances, and to track references. It is thus not surprising
that numerous investigators have considered these fundamental control sys-
tem problems, employing a variety of mathematical formulations and tech-
niques. The regulation problem without internal stability was considered by
Wonham (1973) and that with internal stability by Wonham and Pearson
(1974) and Francis and Wonham (1975) in a geometric setting. Wonham
(1979) gives a complete exposition, where solutions together with (necessary
and sufficient) solvability conditions, and a procedure for the construction of
a compensator have been obtained. This problem was alternatively studied
356 10. Asymptotic Disturbance Decoupling

by Davison (1975) in an algebraic setting. Both works are carried out in a

state-space viewpoint. The problem was also addressed from an input-output
viewpoint by various researchers. Among these are Bengtsson (1977), Francis
(1977), Seaks and Murray (1981) who considered the case where feedback
signals coincide with the outputs; Chen and Pearson (1978), Wolovich and
Ferreira (1979) who considered the case where feedback signals are certain
functions of outputs; and finally Pernebo (1981), Cheng and Pearson (1981)
and Khargonekar and Ozguler (1984) who considered the general case where
feedback signals may be different from outputs and not related to each other.
The crucial condition for the solvability in frequency domain setting is the
existence condition for the solution of Diophantine equation (Kucera, 1979),
which, in a special case, is simplified to skew-primeness (Wolovich, 1978).
The case considered in this chapter is more general than that considered
by Bengtsson (1977) who assume Gm = I and by Wolovich and Ferreira
(1979) who assume Gd (s) = P −1 (s)D−1 (s) where both P (s) and D(s) are
polynomial matrices and P (s) is the denominator of left coprime fraction of
plant Gp (s). Section 1 of this chapter summarizes these elegant results from
a polynomial matrix viewpoint.
For periodic disturbance rejection, Hara et al. (1988) pioneered a method
called repetitive control. However, this method potentially makes the closed-
loop system prone to instability, because the internal positive feedback loop
that generates a periodic signal reduces the stability margin. Moon et al.
(1998) proposed a repetitive controller design method for periodic disturbance
rejection with uncertain plant coefficients. It is noted that such a method is
robust at the expense of performance deterioration compared with the nom-
inal case, and further, the disturbance cannot be fully compensated for even
under the ideal case. Another way to solve the periodic disturbance problem
is to use the double controller scheme (Tian and Gao, 1998). However, the
complexity and the lack of tuning rules hinder its application. A plug-in adap-
tive controller (Hu and Tomizuka, 1993; Miyamoto et al., 1999) is proposed
to reject periodic disturbances. An appealing feature is that turning on or off
the plug module will not affect the original structure. However, the shortcom-
ing of this method is that the analysis and implementations are somewhat
more complex than the conventional model based algorithm. Section 2 of this
chapter presents the virtual feedforward control scheme, which is from Zhang
(2000) and has several advantages over others mentioned above.