Models of Linear Systems

1
Models of Linear Systems

Linear systems are usually mathematically described in one of two domains:
time-domain and frequency-domain. The frequency-domain approach (s- or -
domain) usually results in a system representation in the form of a transfer
function. Transfer functions represent the ratio of a system’s frequency-domain
output to the frequency-domain input, assuming that the initial conditions on the
system are zero. Such descriptions are the subject of many texts in signals and
systems.
In time-domain, the system’s representation retains the form of a differential
equation. However, as any student of engineering will attest, differential
equations can be difficult to analyze. The mathematics gets more burdensome as
the order of the equations increases, and the combination of several differential
equations into one single system can be difficult.
In this chapter, we will introduce a time-domain representation of systems
that alleviates some of the problems of working with single, high-order
differential equations. We will describe a system with state variables, which
collectively can be thought of as a vector. Using the language of vector analysis,
we will demonstrate that state variables provide a convenient time-domain
representation that is essentially the same for systems of all order. Furthermore,
state variable descriptions do not assume zero initial conditions, and allow for the
analysis and design of system characteristics that are not possible with frequency-
domain representations. We will begin with some elementary definitions and a
review of mathematical concepts. We will give a number of examples of state
variable descriptions and introduce several of their important properties.
1.1 Linear Systems and State Equations

To define what we mean by a linear system, we will categorize the types of
systems encountered in nature. First, a system is simply the mathematical
description of a relationship between externally supplied quantities (i.e., those
4 Part I. Mathematical Introduction to State Space
coming from outside the system) and the dependent quantities that result from the
action or effect on those external quantities. We use the term “input” or u to refer
to the independent quantity, although we indeed may have no control over it at
all. It merely represents an excitation for the system. The response of the system
will be referred to as the output y. These input and output signals may be constant,
defined as functions over continuous-time or discrete-time, and may be either
deterministic or stochastic. The system that relates the two may be defined in
many ways, so for the time being, we depict it as in Figure 1.1, simply a block
that performs some mathematical operation.
Input Output
System
u y
Figure 1.1 Elementary representation of a system acting on an input and producing

an output.
1.1.1 Definitions and Review

In this section, we present some definitions for systems that will be useful in
subsequent chapters. It is expected that the reader already has some familiarity
and practice with these concepts from signals and systems studies.
Memory: A system with memory is one whose output depends

on itself from an earlier point in time. A system whose output
depends only on the current time and the current input is
memoryless. (1.1)
Systems with memory most often occur as differential equations (continuous-

time), or as difference equations (discrete-time) because closed-form solutions of
such systems require integration (or summation) of a quantity over past time.
Systems with hysteresis are also examples of systems with memory because the
portion of curve on which they operate depends on the past state and the direction
of change of the input. For our purposes, we will have systems we call algebraic,
which are memoryless, and differential or difference, which represent differential
equations or difference equations. Furthermore, our treatment of algebraic
systems will serve as a tool for the more detailed discussion of differential systems
in the latter chapters.
Causality: A system is said to be causal if the value of the

output at time t 0 depends on the values of the input and output
Chapter 1. Models of Linear Systems 5
for all t up to time t 0 but no further, i.e., only for t  t 0

. (1.2)
Systems that are not causal are sometimes called anticipatory, because they
violate the seemingly impossible condition that they can anticipate future values
of a signal, predicting it at some future time for use at the current time.
Anticipatory systems are often used in data filtering and image processing
applications, wherein an entire data set is first acquired, then processed in batch
mode. In such situations, the “next data” is already available for processing at any
given time.
It is known that for a system to be causal, its transfer function (if it has one)
must be proper. That is, the degree of its numerator polynomial must be no greater
than its denominator polynomial. This is true in both continuous-time systems (s-
domain) and discrete-time systems (z-domain).
Time Invariance: Given an initial time t 0 , the output of a

system will in general depend on the current time as well as this
initial time, y  y ( t , t0 ) . A time-invariant system is one whose
output depends only on the difference between the initial time
and the current time, y  y ( t  t0 ) . Otherwise, the system is
time-varying. (1.3)
Time-varying systems are typically systems in which time appears as an

explicit variable in the differential, difference, or algebraic equation that describes
the system. Thus, a time-invariant differential equation must, by necessity, be one
with constant coefficients. Time-varying systems have outputs that depend, in a
sense, on the actual “clock” time at which they were “turned on.” Time-invariant
systems have outputs that depend on time only to the extent that they depend on
how long it has been since they were “turned on.” Thus if the input were shifted
in time, the output would be simply shifted in time as well. Time-varying
equations are very difficult to solve, rivaling nonlinear equations.
To define linearity, we consider the action of the system to be represented by
the symbol S, i.e., using our previous notation, y  S ( u ) . If we consider two
inputs, u1 and u2 , and a scaling factor, a, we introduce the definition:
Linearity: A linear system is one that satisfies homogeneity

and additivity. A homogeneous system is one for which
S ( au )  aS ( u ) for all a and u , and an additive system is one
for which S ( u1  u2 )  S ( u1 )  S ( u2 ) for all u1 and u2 . (1.4)
Linear systems are thus systems for which the principle of superposition
holds. We will later consider so-called multivariable systems, which have more
than one input and more than one output. When such systems are linear, the effect
of each input can be considered independently of one another. In systems with
memory, the term linear refers to systems that are linear in all of the variables on
which they depend. Therefore, for example, a linear nth order differential equation
is one whose nth derivative depends in a linear way on each of the lower
derivatives, and also in a linear way on the forcing function, if any.
Nonlinear systems are notoriously difficult to analyze and solve, partly
because they exist in such an infinite variety of forms, preventing any cohesive
theory for analysis.
In the next section, we will review the process by which models of linear
systems are derived, followed by some examples for practical physical systems.
1.1.2 Physical System Modeling

The underlying motivation for all the analysis tools presented in this book is the
understanding of physical systems. Whether the system is mechanical, electrical,
or chemical, a mathematical description must be written in a unified way so that
a single theory of stability, control, or analysis can be applied to the model. This
is often the first task of an engineer in any design problem. In this section, we will
introduce linear modeling principles for electrical, mechanical, and some fluid
systems, and we will attempt to illustrate the unity in such models.
Physical Variables
We start by categorizing the physical quantities of interest to us. The first
quantities available in any problem specification are the constants. These are, of
course, constant numerical values specifying the dimensions, ranges, amounts,
and other physical attributes of the masses in the system. These are often available
as known quantities, but are sometimes unknown or poorly known and are subject
to a process known as system identification. System identification is also useful
when the physical attributes of a system are not constant, but vary with time. For
example, the weight of a vehicle may change as its fuel is consumed, and the
resistance of a resistor may change with temperature. In this chapter and for most
of this book, we will not consider time-varying quantities in much detail.
The second class of number to be considered is the variables, which are of
interest to us because they do usually vary with time. Some variables, i.e., those
considered inputs, are known a priori, while others (the outputs) are to be
determined. We separate these into two broad categories: flow variables and
potential variables. Flow variables are quantities that must be measured through
the cross-section of the medium through which they are transmitted. The easiest
flow variable to imagine is a fluid (nonviscous, incompressible). The flow of such
a fluid through a conduit must be measured by breaking the pipe and “counting”
the amount (mass) of fluid passing through the cross-section.* For electrical
systems, the analogous flow variable can be considered to be the current. Because
current is defined as the amount of charge per unit area flowing across a cross-
section per unit time, we can equally well consider charge to be a flow variable.
In mechanical systems, force is a flow variable. Although it may not conform to
the fluid analogy, it is nevertheless a quantity that must be measured by breaking
a connection and inserting a measurement device.
The second type of variable is the potential variable. Potential variables are
physical quantities that must be measured at two locations; the value of the
measurement is the relative difference between the locations. Pressure, voltage,
and displacement (position or velocity) are all potential variables because their
definitions all require a reference location. Although we speak of a voltage
appearing at a particular location in a circuit, it is always understood that this
measurement was taken relative to another point.
Physical Laws
For many simple systems, there are only a few basic physical laws that must be
obeyed in order to generate a sufficient set of equations that describe a system. If
we consider only the basic necessities (i.e., using finite-dimensional, lumped-
parameter, Newtonian dynamics rather than relativistic mechanics), we can
categorize these into two types of laws: mass conservation and circuit
conservation laws. Mass conservation laws are defined on nodes, and circuit
conservation laws are defined on closed paths. A node is an interconnection point
between two or more conduits transmitting a flow variable. In mechanical
systems, nodes are associated with masses so that applied forces are shown to be
acting on something. These circuit laws are integrally related to the two types of
variables above: flow and potential.
Mass conservation laws take the basic form:
F all flow variablesIJ   FG net equivalent flowIJ

 GH entering a node K H into node K (1.5)
For an electrical network, this type of law translates to Kirchoff’s current law
(KCL), which states that the sum of all currents entering a node must equal zero,
 ii  0 . For a mechanical system, the mass conservation law takes the form of
Newton’s law:  Fi  ma . Note that in Newton’s law, the sum of flow variables
need not equal zero but must be proportional to the net acceleration of the object
on which the forces act. In electrical and fluid systems, the net equivalent flow is
* In reality, clever fluid flow measurement systems have been devised that can measure
flow variables without interrupting them, e.g., ultrasonic flowmeters and inductive
ammeters.
zero, because it is impossible for net charges to accumulate indefinitely in a wire

connection, just as it would be for fluid molecules to accumulate in a conduit
junction.
Circuit conservation laws take the following form:
F signed changes in a potential I

 GH variable around a closed pathJK  0 (1.6)
Such laws enforce the intuitive notion that if a potential variable is measured at
one location relative to a fixed reference, and if relative changes are added as
components are traversed in a closed path, then the potential measured at the
original location upon returning should not have changed. Thus, Kirchoff’s
voltage law (KVL) specifies that around any closed path in a network,  v i  0 ,
being careful to include the appropriate algebraic signs in the terms. In mechanical
systems, circuit conservation allows us to measure what we consider absolute
position by summing a sequence of relative displacements (although in truth all
positions are relative to something). For fluids, the sum of pressure drops and
rises throughout a closed network of pipes and components must equal zero.
These laws go a long way toward generating the equations that describe
physical systems. We are ignoring a great many physical quantities and processes,
such as deformation, compressibility, and distributed parameters, that usually
provide a more complete and accurate model of a system. Usually, though, it is
best to attempt a simple model that will suffice until its deficiencies can be
discovered later.
Constitutive Relationships
The physical laws above are not by themselves sufficient to write complete
equations. Flow variables and potential variables are not unrelated, but their
relationship depends on the physical device being considered. Aside from
sources, which provide, e.g., input forces, voltages, currents, and flows to
systems, we also have components, which we assume to be lumped, i.e., their
effects are modeled as being concentrated at a single location as opposed to being
distributed over space. Each type of component has associated with it a
constitutive relationship that relates the flow variable through it and the potential
variable across it.
Electrical Components
The three most basic linear components common in electrical networks are the
resistor (R, measured in ohms,  ), capacitor (C, measured in farads, F), and
inductor (L, measured in henrys, H). These are pictured in Figure 1.2.
iR iL iC
R vR L vL C vC
Figure 1.2 Electrical components: resistor, inductor, and capacitor (left to right).
For each component, we define reference directions for the flow variable
(indicated by an arrow) and the potential variable (indicated by  /  signs), so
that the appropriate algebraic sign can be defined for the component.
For these three components, the constitutive relationships are:
diL dvC
v R  iR R vL  L iC  C (1.7)
dt dt
At any such component, these relationships can be used to substitute a flow

variable for a potential variable or vice versa. Note that with the differential
relationships, an integration is necessary when expressing the reverse
relationship.
Mechanical Components
For mechanical systems, the fundamental components are the mass, damper, and
spring. These are pictured in Figure 1.3 below.
FM FK FB
M
xM K xK B xB
Figure 1.3 Mechanical components: mass, spring, and damper (left to right).
In the figure, a displacement is shown which indicates that one side of the
component is displaced relative to the other side, except for the mass, which is
displaced relative to a fixed reference defined elsewhere. The constitutive

equations for these elements are
d 2x M dx B
FM  M FK  Kx K FB  B (1.8)
dt 2 dt
where F is the force applied to (or resulting from) the component. It should be
noted that these equations do not appear entirely analogous to those of the
electrical system (1.7). Most notably, these equations have a second derivative,
and are all expressions of the flow variables in terms of the potential variable.
This is partly based on convention so that the requisite computations are more
convenient, and partly natural, since electrical and mechanical quantities are
perceived differently by the systems on which they act. Nevertheless, the
mathematical analogies remain. For fluid systems, the analogy becomes even
weaker, with tanks and valves being the primary components. These components
have somewhat more complex, sometimes nonlinear constitutive relationships. In
truth, though, all constitutive relationships become nonlinear when the limits of
their capacity are approached.
Example 1.1: Mechanical System Equations

Derive the equations of motion for the system of two masses shown in Figure 1.4.
In the system, the two masses are connected by a spring with Hooke’s law
constant K and a damper with damping constant B, both initially unstretched and
stationary. Their positions on the horizontal plane are measured as x1 ( t ) and
x 2 ( t ) from points on the masses such that they are initially equal. In addition, an
externally applied force F pushes on the first mass.
x1
F K
m1
m2
B
x2
Figure 1.4 Two-mass system containing a spring and damper, and a forcing
function.
Solution:
As mass 1 moves in the direction of positive x1 , the spring will compress and
react with a force K ( x1  x 2 ) against the motion. Likewise, the damper will resist
motion with viscous friction force B( x1  x 2 ) . The free-body diagram of the
system in Figure 1.5 shows the two masses and all of the forces acting on them.
K ( x1  x 2 )
F m1
B ( x1  x 2 )
K ( x1  x 2 )
m2
B ( x1  x 2 )
Figure 1.5 Free-body diagrams showing the masses in Figure 1.4 and the forces that
act on them.
Applying Newton’s law, ma   F , we get
b g b
x1  F  K x1  x2  B x1  x 2
m1 g (1.9)
For the second mass, only the spring and damper provide forces, which are equal
and opposite to the forces seen in (1.9). Therefore,
b g b
x2  K x1  x2  B x1  x 2
m2  g (1.10)
Rearranging these equations to a more convenient form,
b g b
x1  B x1  x 2  K x1  x2  F
m1 g
x  Bb x  x g  K b x x g0
(1.11)
m 
2 2 2 1 2 1
These simple linear equations will be used for further examples later.
Example 1.2: Electrical System Equations

For the circuit shown in Figure 1.6, derive differential equations in terms of the
capacitor voltage v c ( t ) and the inductor current i( t ) .
vx
R1
vR
C
ic  i
R2
v vC L vL
Figure 1.6 Electric circuit example with variables v c ( t ) and i( t ) .
Solution:
In this example, the sum of the three currents entering node 1 must be zero. While
Ohm’s law v  iR may be used for the resistor, the current through the capacitor
is given by ic  C dvc dt . Thus,
vc dv
 C c i  0 (1.12)
R1 dt
where care has been taken to use only the desired variables v c ( t ) and i( t ) . For a
second relationship, we note that the sum of all voltages around the main loop of
the circuit must be zero. The voltage across the inductor is given by v  L di dt ,
so
di
 v  vc  R2i  L 0 (1.13)
dt
where v ( t ) is the forcing function provided by the voltage source in the circuit.
Rewriting (1.12) and (1.13) in a more convenient form,
dvc 1 di
C  vc  i  0 L  R2i  vc  v (1.14)
dt R1 dt
1.1.3 State Equations

The sets of equations derived in (1.11) and (1.14) are coupled, in the sense that
the variables in one appear in the other. This implies that they must be solved
simultaneously, or else they must be combined into a single, larger-order
differential equation by taking derivatives and substituting one into the other. The
standard methods for solving differential equations are then applied.
However, such a process can be tedious, and the methods employed vary in
complexity as the order of the differential equation increases. Instead, we prefer
to write the dynamic equations of physical systems as state equations. State
equations are simply collections of first-order differential equations that together
represent exactly the same information as the original larger differential equation.
Of course, with an nth-order differential equation, we will need n first-order
equations. However, the variables used to write these n first-order equations are
not unique. These so-called “state variables” may be chosen for convenience, as
one set of state variables may result in mathematical expressions that make the
solution or other characteristic of the system more apparent.
In a strict sense, the collection of state variables at any given time is known
as the state of the system, and the set of all values that can be taken on by the state
is known as the state space. The state of the system represents complete
information of the system, such that if we know the state at time t 0 , it is possible
to compute the state at all future times. We will model the state spaces for linear
systems as linear vector spaces, which we begin to discuss in the next chapter.
State Variables
Consider a nth-order linear, time-invariant differential equation:
d n x( t ) d n 1x( t ) dx( t )
n
 a n 1 n 1
   a1  a0 x( t )  b0u( t ) (1.15)
dt dt dt
The most straightforward method for choosing n state variables to represent this
system is to let the state variables be equal to x ( t ) and its first ( n 1) derivatives.
Thus, if the state variables are denoted by  , then
1 ( t )  x ( t )
dx ( t )
2 (t ) 
dt

d n 1 x ( t )
 n (t ) 
dt n 1
These definitions of state variables are also called phase variables. The n
differential equations resulting from these definitions become
 1 ( t )   2 ( t )
 ( t )   ( t )
2 3
 (1.16)
 n 1 ( t )   n ( t )
 n ( t )   a0 1 ( t )  a1 2 ( t )    a n  2 n 1  a n 1 n  b0u( t )
We will find it convenient to express such a system of equations in vector-

matrix form:
LM  OP LM 00 1 0  0 OP L  O L 0 O
MM PP  MM 0
1
0 1  0 PP MM PP MM  PP
1
PP MM  PP  MM 0 PPu(t )
2
2

MM PP MM 0
0 0 0 (1.17)
0  0 1
N Q MNa
n
0  a1  a2   a n 1 PQ N Q Nb Q
n 0
If in addition, one of the state variables, say 1 ( t ) , is designated as the “output”

of interest, denoted y ( t ) , then we can also write the so-called “output equation”
in vector-matrix form as well:
LM  OP
1
0 M P

y( t )  1 0 
MM  PP
2
(1.18)
N Q
n
More generally, we may designate as the output a weighted sum of the state
variables and sometimes also a sum of state variables and input variables.
 
Defining x( t )  1   n , the two equations (1.17) and (1.18) are together
M
written as
x  Ax  bu
(1.19)
y  cx  du
ss(a,b,c,d)
ssdata(sys) where the matrices A, b, c, and d are the corresponding matricesM in (1.17) and
(1.18). These will be referred to as the state matrix (A), the input matrix (b), the
output matrix (c), and the feedthrough matrix (d) (so named because it is the gain
through which inputs feed directly into outputs). These equations are expressed
for a single input, single output (SISO) system. For multi-input, multioutput
(MIMO) or multivariable systems, the equations in (1.19) can be written exactly
the same except that the input matrix is a capital B, the output matrix is a capital
C, and the feedthrough matrix is a capital D. These changes indicate that they are
matrices rather than simple columns (b), rows (c), or scalars (d). The equations in
(1.19) will become quite familiar, as they are the format used for studying the
properties of linear systems of equations throughout this book.
Of course, if the original Equation (1.15) were time-varying, then the
coefficients might be functions of time, i.e., ai ( t ) and bi ( t ) . In that case, (1.19)
might contain A( t ), b( t ), c( t ) and d ( t ) .
Example 1.3: State Variables for the Mechanical System Example

Write a state variable expression for the differential equations already derived for
the mechanical system of Example 1.1, using force F as the input and the
difference x 2  x1 as the output.
Solution:
In the mechanical system, we derived two separate equations, each being second
order (1.11). To generate state equations, we will introduce the variables
1  x1 ,  2  x1 ,  3  x2 , and  4  x 2 . Then, by inspection of (1.11), the state
equations are:
LM  OP LM K0 1 0 0 OP L  O L 0 O
PP MM PP  MM 1 m PP F (t )
1 1
MM PP  MM 0 m  B
m1
K
m1
B
m1
PP MM PP MM 0 PP
2 1 2
1
MM PP MM K
3
0 0 1 3
N Q N m K m  Bm
Q N Q MN 0 PQ
B
4 2 m2 2 2 4
(1.20)
LM  OP 1
0 M P

y ( t )  1 0 1
MM PP
2
N Q 4
Example 1.4: State Variables for the Electrical System Example

Find a state variable expression for the electrical system equations in (1.14) from
Example 1.2. As the system output, use the voltage across the inductor.
Solution:
In this system, the two equations in (1.14) are each first order. The total system is
then second order. Using v c and i( t ) as the two state variables, i.e.,
1 ( t )  vc ( t ) and  2 ( t )  i( t ) , we can immediately write the equations:
LM OP  LM 1CR 1 OP L  O L 0 O

MN PQ MN  1 L
1 1 C
R2 M P  M 1 P v bt g
 L P N Q MN L PQ
1
1
Q 2
(1.21)
y ( t )  1  R2
LM  OP  b1gv(t )
1
N Q
2
The output equation in this result comes from the KVL equation v L  v  v c  iR2
.
Alternative State Variables

If we had attempted to use the definitions of state variables in (1.16) to write state
equations for the more general differential equation:
d n x( t ) d n 1x( t ) dx( t )
n
 a n 1 n 1
   a1  a0 x ( t )
dt dt dt
(1.22)
d n u( t )
 bn    b0u( t )
dt n
we would have required derivatives of u( t ) in the state equations. According to

the standard format of (1.19), this is not allowed. Instead, there are a number of
commonly used formulations for state variables for equations such as (1.22).
These are best represented in the simulation diagrams on the following pages.
In Figure 1.7, the state variables shown are similar to those found in (1.16),
in the sense that n 1 of them are simply derivatives of the previous ones. The
state equations that describe this diagram are given immediately following Figure
1.7, in Equation (1.25). We see that in this form, the feedback coefficients a i
appear in only the final state equation. Having feedforward connections with bi
coefficients allows for derivatives of u( t ) in (1.22) without appearing in (1.25)
below, which is the state variable equivalent of (1.22). Note that in this form each
state variable is assigned to be the output of an integrator, just as with the phase
variables discussed above.
A second common choice of state variables can be generated from the
following manipulations. Suppose that in an attempt to solve the second-order
equation of the form:
x ( t )  a1 x ( t )  a0 x ( t )  b2 u( t )  b1u( t )  b0 u( t ) (1.23)
both sides of the equation are integrated twice with respect to the time variable.
This would result in:
zz
t s
x (  )  a1x (  )  a0 (  ) x(  ) d ds
 
zz
t s
 b2u(  )  b1u(  )  b0 (  )u(  ) d ds
 
giving
z zz
t t s
x ( t )  a1x( s ) ds  a0 x (  ) d ds
  
z zz
t t s
 b2u( t )  b1u( s ) ds  b0u(  ) d ds
  
or
z zz
t t s
x ( t )  b2 u( t )  b1u( s )  a1 x ( s ) ds  b0 u(  )  a 0 x (  ) d ds
  
z R|ST| U| (1.24)
z
t s
 b2 u( t )  b1u( s )  a1 x ( s )  b u(  )  a x (  ) d V ds
 
0 0
W|
For higher-order systems, this process continues until an equation of the form of
(1.24) is derived. From the form of (1.24), the simulation diagram shown in Figure
1.8 can be drawn, with the associated state equations appearing in Equation
(1.26). Note that in this formulation, the integrators in the network are not
connected end-to-end. Thus the state variables are not simply derivatives of one
another as are phase variables. Instead, the state equations are written as in (1.26).
The state variables are, however, still defined as the outputs of the integrators.
This is commonly done, but is not necessary. Additional examples will be shown
in Chapter 9.
bn
bn 1
b1
u (t ) xn x n 1 x2 x1
-
+
xn
 
x n 1 x1
 b0 + y (t )
- -
a n 1
an  2
a0
LM x 1 OP LM 0 1 0 0 OP LM x 1 OP LM0OP
MM   PP  MM    0 PP MM  PP  MM  PPu(t )

MNxx
n 1
n
PQ MN0a 0
0
 a1   a n 1
1
PQ MNxx
n 1
n
PQ MN10PQ
LM x 1 OP (1.25)
y ( t )  b0  bn a 0 b1  bn a1  bn 1  bn a n 1 MM  PP  b u(t )
n
MNxx
n 1
n
PQ
Figure 1.7 Simulation diagram and state equations for phase variable definitions of
state variables.
These special choices are examples of different ways in which state variables
can be assigned to a particular system. They have some convenient properties that
we will examine in later chapters. Such special forms of equations are known as
canonical forms.
It should be noted that there are an infinite number of other ways in which to
derive definitions for state variables, a fact that will soon become readily
apparent. It is also important to realize that in our simple physical system
examples, the variables we choose are physically measurable or are otherwise
meaningful quantities, such as voltages or displacements. Often, variables will be
selected purely for the effect they have on the structure of the state equations, not
for the physical meaning they represent. State variables need not be physically
meaningful. This is one of the primary advantages to the state variable technique.
bn
b1
xn + x n 1 x1 +
y (t )
  
+ + +
b0
u (t )  xn  x n 1 x1
a n 1
a1
a0
LM x 1 OP LMa n 1 1 OP LM x OP LMb  a b OP
0 0 1 n 1 n 1 n
MM   PP  MM    0 PP MM  PP  MM  PPu(t )
MNxx
n 1
n
PQ MN aa
1
0
 0 1 x
P M P
0  0Q N x Q N b  a b Q
M b a b
P n 1
n
1
0
1 n
0 n
LM x OP 1
(1.26)
 0 M
 P
y(t )  1 0
MMx PP  b u(t ) n 1
n
Nx Q n
Figure 1.8 A second type of simulation diagram and state equations.
Changing State Variables

In the state variable description for the mechanical system, we might be inclined
to write equations not for the independent variables, x1 ( t ) and x 2 ( t ) , but rather
b g
for the variables x1 ( t ) and x1 ( t )  x2 ( t ) . Perhaps we make this choice because
the device is being used as a seismograph and the motion of the second mass
relative to the first is more important than the absolute position. For this system,
the equations in (1.11) are relatively simple to rewrite with this change of
variables. If we introduce x1 ( t )  x1 ( t ) and x 2 ( t )  x1 ( t )  x2 ( t ) , then (1.11)
becomes:
m1x1  Bx2  Kx2  F

(1.27)
m2  e
x1   j
x2  Bx2  Kx2  0
In order to get each equation to contain second derivatives in only one of the
variables, we can solve the first equation of (1.27) for 

x1 and substitute it into the
second equation. Performing this operation and simplifying, we obtain
m1 
x1  Bx 2  Kx 2  F
m2 
FG m IJ m FGm
x 2  B 1  2 x 2  K 1  2 x 2  2 F
IJ
H m1 K m1 Hm1 K
Now using new state variables  1  x1 ,  2  x1 ,  3  x 2 and  4  x 2 , the
new state equations are readily apparent:
LM  OP LM0 1 0 OP L  O L 0 O
0
MM  PP MM0 PP MM  PP MM 1 PP
1 1
K B
PP MM PP  MM m PP F
0  
MM PP  MM0
2 m1 m 1 2 1
MM PP MM 0 0 1
P MM PP MM 0 PP
3
F 1  1 IJ  BFG 1  1 IJ PP M  P MM 1 PP 3
MN PQ MN0
4
0 KG
H m m K H m m K Q MN PQ N m Q
1 2 1 2 4 1
LM  OP
1
MM PP
1 0 M P
2
y 0 0
MM PP
MM  PP
3
N Q
4
Here again we have let the output variable be y ( t )  x 2  x1 (   x 2 ( t )    3 ) .

Notice that the state equations above are significantly different from (1.20),
although they are both valid equations representing the same physical system.
Clearly, redefining all the state variables and rederiving or manipulating the
equations can be time consuming. To demonstrate an easier method that uses the
vector-matrix notation we have adopted, we turn to the electrical system example
(Example 1.2). Suppose we wish to write new state equations, where the relevant
state variables are v R and v x (see the circuit diagram in Figure 1.6). Because we
can write the new state variables as weighted sums of the old state variables
v R  R2i
(1.28)
v x  vc  R2i
we can use vector-matrix notation immediately to write
LMv R OP LM0 R2OP LMv OP c
MN v x
PQ  MN1 PM P
R QN i Q
2
(1.29)
Alternatively, using the inverse relationship,
LMv OP  LM0 R2OP LMv OP  LM R1  0R OP LMv OP  LM 11

1 2 2 OP L O
1 vR
R Q Nv Q N Q N v Q MN R PQ MN PQ
c R R
(1.30)
N i Q N1 2 x
R 2
x 2
0 vx
  
If we use the notation   v C i  1 2 for the state vector already
  
  v
introduced in (1.21) and  R vx   1  2 for the new state vector,
then we can write (1.30) symbolically as

  M (1.31)
where matrix M is defined in (1.30). Likewise, the symbolic form for (1.21) is as
given in (1.19), i.e.,
  A  bv

(1.32)
y  c  dv

  M
Equation (1.31) directly implies that  , so substituting into (1.32),
   bv
M  AM
(1.33)
y  cM   dv
or
   M 1bv  A  
  bv
  M 1 AM
(1.34)
  dv  c
y  cM 
  dv
where
LM0 R OP LM CR1 OP L 1

1 1OP
MN1 R PQ MM  1 PP MM 1
2

A  M 1
AM  1 C
P
QP MN R
R 0P
2
MN L  2
L 2 Q
LM 0 R
 2
OP
M P
L
MM 1  1 
1

R P
L PQ
2
N CR CR
1 2 CR1
L 0 R OL 0 O M P
LR O 2
b  M b  M M P L
MN1 R PPQ MMN 1 LPPQ  MM R PP
2
 1
MN L PQ
2
2
LM 1 1OP
c  cM  1  R M 1 P

 0 1 (1.35)
MN R PQ
2
0
2

d  d  1
This gives a new system of state equations that are equivalent to those in (1.21)
but appear quite different. Notice that the feedthrough matrix, d, is the same in
each case.
The procedure we have just introduced will be considered from a different
perspective in later chapters. The important points here are:
• nth-order linear differential equations (or several coupled systems
constituting an nth-order system) can be written as n first-order state
equations.
• State equations are entirely equivalent to the original differential
equations but are not unique.
• State equations can be changed with matrix-vector operations, resulting
in a new form for each of the four system matrices.
In future chapters, we will see what conveniences can be realized with
different choices of state vector definition and what insight might be gained into
the original physical system by using the state variable notation. There will be
certain characteristics of differential equations and control systems that are more
apparent in state space (time-domain) than in transfer function form (frequency-
domain). In order to understand these properties, some details of linear vector
spaces and linear algebras will be necessary.
1.1.4 Discrete-Time Systems

Throughout this book we will, on occasion, use discrete-time systems as examples
or for other illustrative purposes. Perhaps more so than in frequency-domain, state
space methods for discrete-time systems are very similar to those used in
continuous-time systems. For example, most of the first part of the book on vector
spaces, linear operators, and functions of matrices is indifferent to the time-
domain, because the equations being considered are independent of time. In latter
chapters, some concepts, such as controllability and pole placement are common
to continuous and discrete-time, while other concepts, including stability and
advanced controller design, can vary significantly in discrete-time. The basic
concepts remain parallel, but the matrix equations may look different.
We do not give a detailed treatment here of digital filtering or z-domain
methods, or of continuous filtering and s-domain methods. Nevertheless, it is
useful to review the terminology.
Discrete-time systems may be inherently discrete, as, for example, in the
equations that describe the balance of a bank account that undergoes withdrawals,
deposits, and interest postings at regular intervals. Alternatively, they may be
discretizations of continuous-time systems. A discretization is a conversion of a
continuous-time equation into discrete-time. These discretizations may be
performed in a number of different ways, e.g., by using integrator equivalence,
pole-zero mapping, or hold equivalence [6]. In either case, we end up with a
difference equation rather than a differential equation. Difference equations are
expressed in terms of time delays rather than derivatives. If, for example, the
sampling period of a discrete-time system is T, then a simple difference equation
might appear as
x ( kT  2T )  a1 x ( kT  T )  a0 x ( kT )  b0 u( kT ) (1.36)
where k is an integer. In such a situation, we will sometimes simplify the notation

by dropping the (constant) sampling time T and using a subscript rather than an
argument in parentheses, i.e.,
x k  2  a1 x k 1  a0 x k  b0 uk (1.37)
With this difference equation, we may define a state vector in any of the ways
discussed above for continuous-time systems. For example, if we let
 
  xk x k 1 , then we will obtain the discrete-time state equation
 k 1 
LM 0 1 OP  LM 0 OPu
N a 0  a1 Q Nb Q
k
0
k
(1.38)

 Ad  k  bd uk
where the notation Ad and bd are introduced to distinguish these discrete-time

system matrices from their continuous-time counterparts. An output equation can
be similarly defined.
Furthermore, if an arbitrary difference equation is specified:
x k  n  a n 1 x k ( n 1)    a1 x k 1  a0 x k  bn uk  n    b0 uk (1.39)
then the simulation diagrams of Figures 1.7 and 1.8 (and all other possibilities)
may also be used, with the exception that the integrators are replaced with unit-
time delays, i.e., delay blocks representing one sample period, T, in length. When
such a system is time-varying, the coefficients a i and bi above may be replaced
with ai ( k ) and bi ( k ) . (The inclusion of this extra time argument is the
motivation for the subscript notation; writing ai ( k , kT  T ) would be
cumbersome.)
The following examples illustrate a number of ways a difference equation
can be used to model physical systems. The first two examples illustrate the
derivation of difference equations by direct discrete-time modeling, and in the
third example, a method is given for approximating a differential equation by a
discrete-time system. Another method for representing discrete-time systems as
approximations of continuous-time systems is given in Chapter 6.
Example 1.5: Direct Difference Equation Modeling of a Savings Account

The simplest model of an interest-bearing savings account provides an example
of a first-order difference equation. Derive the difference equation model for the
balance of an account earning i% per year, compounded monthly. Assume that
interest is computed on the previous month’s balance and that the account owner
may make any number of deposits and withdrawals during that month.
Solution:
The balance, after compounding, in month k is denoted x ( k ) . Then x ( k ) will be
equal to the previous month’s balance plus the interest for month k and the net
total of the owner’s deposits and withdrawals, which we will denote by u( k ) .
Then we have the difference equation
x ( k )  x ( k  1)  i 12 x ( k  1)  u( k  1)
e j
 1  i 12 x ( k  1)  u( k  1)
This is a first-order difference equation and, hence, is already in linear state

variable form. Note that time can be arbitrarily shifted to reflect that k  0 is the
time origin and that x(1) is the first balance that must be computed. Because the
system is time-invariant, this is merely a notational convenience:
e j
x ( k  1)  1  i 12 x ( k )  u( k )
Example 1.6: A Difference Equation for a Predator-Prey System

A tropical fish enthusiast buys a new fish tank and starts his collection with Pp ( 0 )
piranhas and Pg ( 0 ) guppies. He buys an ample supply of guppy food (which the
piranhas will not eat) and expects the piranhas to eat the guppies. He samples the
populations of the two species each day, Pp ( d ) and Pg ( d ) , and finds, as he
expected, that they change. Generate a linear difference equation model for these
“population dynamics.”
Solution:
Several assumptions must be made to derive a suitable model. First, we will
assume that the birthrate of the piranhas is directly proportional to their food
supply, which of course is the population of guppies. Further, we assume that
because piranhas are large, their death rate is proportional to overcrowding, i.e.,
to the level of their own population. Therefore, we can write the relationship
Pp ( d  1)  Pp ( d )  k1Pg ( d )  k2 Pp ( d )
b g
 1  k2 Pp ( d )  k1Pg ( d )
(1.40)
where k1 and k 2 are constants of proportionality. Of course, Pp ( 0 ) and Pg ( 0 )

are the initial conditions.
Now assume that the birthrate of the guppies is proportional to their food
supply, which is the input of guppy food, u( d ) . The death rate of the guppies will
be proportional to the population of the piranhas. Therefore,
Pg ( d  1)  Pg ( d )  u( d )  k3 Pp ( d ) (1.41)
Together, these two equations can be combined into the state space difference
equation model
LM P ( d  1)OP L1  k k1OP LM P ( d )OP LM0OP

MN P ( d  1)PQ  MMN  k
p 2 p
g 3
P M P  M Pu( d )
1 Q N P ( d )Q N1Q
g
(1.42)
Example 1.7: Discretized Differential Equations for a Harmonic Oscillator

By approximating the derivative as a finite difference, find a discrete-time version
of the forced harmonic oscillator equations
x ( t )   2 x ( t )  u( t )
 (1.43)
Solution:
At time t  kT , where T is a sampling time, the approximation of the first
derivative is
x ( kT  T )  x ( kT )
x ( kT ) 
T
This implies that the approximation for the second derivative would be:
x ( kT  T )  x ( kT )
x ( kT ) 

T
x ( kT  2T )  x ( kT  T ) FG
x ( kT  T )  x ( kT ) IJ

T

H T K
T
x ( kT  2T )  2 x ( kT  T )  x ( kT )

T2
Substituting these approximations into the original Equation (1.43), we get
x( kT  2T )  2 x( kT  T )  x( kT )
2
  2 x( kT )  u( kT )
T
or simplifying and dropping the T ’s in the time arguments,
e j
x ( k  2)  2 x ( k  1)   2T 2  1 x ( k )  T 2u( k ) (1.44)
This can be seen as a discrete-time equation of the form (1.39).

If we now choose state variables as x1 ( k )  x ( k ) and x 2 ( k )  x ( k  1) , then
the discrete-time state space description of the system becomes
LM x ( k  1) OP LM 0
1 1OP LM x ( k ) OP LM 0 OP
1
MNx ( k  2)PQ  MN( T

2
2 2
 1)
P M P  M Pu( k )
2Q N x ( k )Q NT Q
2
2
(1.45)
If we now wish to draw a simulation diagram for this system that is similar to the
one in Figure 1.7, we must remember that instead of integrators, we must use the
unit delay operator, z 1 . Because (1.45) has the same form as (1.39), which itself
is similar in form to (1.22), the simulation diagram of Figure 1.9 is obtained.
u(k )  x 2 ( k  2) x1 ( k  1)
z 1 z 1 T2 x (k )
x2 (k ) x1 ( k )
 
2
2 T 2  1
Figure 1.9 Simulation diagram for a discrete-time state space system. The unit delay
1
operator is signified by the symbol z .
1.1.5 Relationship to Transfer Functions

It may be reasonably asked at this point how the state variable description, which
is in time-domain, relates to the transfer function representation of systems that is
usually introduced in lower level signals and systems courses. The most
fundamental answer to this question is that transfer functions are defined for
systems with zero initial conditions only. State variable descriptions are therefore
more appropriate when transients due to nonzero initial conditions are important.
However there are many other distinctions as well. The mathematical
procedures for solving time- and frequency-domain systems differ greatly. From
a control systems point of view, there exist entirely different design tools in time
and frequency-domains, and certain aspects of system behavior are more obvious
in one domain or the other. Design criteria can be specified in one domain that
are not easily expressed in the other. Other concepts that will be introduced in
later chapters include controllability, observability, and optimal compensator
design, concepts that are all straightforward in time-domain but not necessarily in
frequency-domain.
For the time being, though, we will give a flavor of the intrinsic relationship
between the two domains. Consider the form of general SISO state equations:
x  Ax  bu
(1.46)
y  cx  du
If we assume that the initial condition is x ( t 0 )  0 for this system, Equations

(1.46) can be LaPlace transformed simply as
sX ( s )  AX ( s )  bU ( s )
(1.47)
Y ( s )  cX ( s )  dU ( s )
Simplifying the first of these equations,
b g
sX ( s )  AX ( s )  sI  A X ( s )  bU ( s )
or
b
X ( s )  sI  A g 1
bU ( s ) (1.48)
Substituting (1.48) into the output equation in (1.47), we get
b
Y ( s )  c sI  A g 1
b  d U ( s)
or, for a SISO system in which Y ( s ) and U ( s ) are scalar quantities and the
division operation exists,
Y ( s)
U ( s)
b
 c sI  A g 1 
b  d  P( s ) (1.49)
giving the transfer function, which is a ratio of polynomials.

It is worth noting that many textbooks pay little attention to the feedthrough
term of the state equations, i.e., d (D). This is because its effect can be “factored
out” such that it represents an algebraic relationship between the input and output,
rather than a differential one. It occurs only in nonstrictly proper transfer functions
wherein the order of the numerator is the same as the order of the denominator.
By using polynomial division, such a transfer function can always be rewritten as
the sum of a constant factor and a strictly proper transfer function. For example,
consider the transfer function
2s  4 2
P( s )   2 (1.50)
s 1 s 1
Converting this system to state space representation will always give d  2 as the
feedthrough term, regardless of the choice of variables. Recall that if the transfer
function is improper, then positive powers of s would divide out of such a fraction,
and a noncausal system would be apparent.
Equation (1.49) shows how it is possible to obtain a transfer function from
the matrices given in a state variable representation.M To get a state variable tf2ss(num,den)
representation from a transfer function,M we usually return to the original ss2tf(a,b,c,d)
differential equation or use the simulation diagrams in Figures 1.7 and 1.8. This
subject is complicated by controllability and observability issues introduced in
Chapters 8 and 9.
We have asserted that no matter which (valid) state variables we choose, the
state equations represent the original system exactly. One may than ask if a
change of variables is performed as in (1.31), how is the transfer function (1.49)
altered? Suppose in the equation for the transfer function in terms of system
matrices in (1.49), we substitute the “transformed” matrices from (1.35). Then we
obtain
1
e
P ( s )  c sI  Aj b  d
1
e
 cM sI  M 1 AM j M 1b  d
1
 cM e sM 1
M  M 1 AM j M 1b  d (1.51)
b
 cMM 1 sI  A g 1
MM 1b  d
 cb sI  Ag
1
b d
 P( s )
So we have shown that the same transfer function results regardless of the choice
of state equations. Note that this same result holds for discrete-time systems in
exactly the same form; simply replace the s operator in this section with the z
operator.
1.2 Linearization of Nonlinear Equations

It is an unfortunate fact that most physical systems encountered in practice are not
linear. It is almost always the case that when one encounters a linear model for a
physical system, it is an idealized or simplified version of a more accurate but
much more complicated nonlinear model. In order to create a linear model from
a nonlinear system, we introduce a linearization method based on the Taylor series
expansion of a function.
1.2.1 Linearizing Functions

Recall that the Taylor series expansion expresses a general function f ( x ) as the
infinite series

1 d n f (x)
f (x)   n! dx n
( x  x0 ) n (1.52)
n0 xx 0
This series is said to be expanded about the point x  x 0 . The point x 0 is

interchangeably referred to as the bias point, operating point, or, depending on
some stability conditions discussed in Chapter 7, the equilibrium point.
Any function for which such a series converges is said to be analytic. Writing
out the first few terms of such a series,
df ( x ) 1 d 2 f (x)
f ( x )  f ( x0 )  ( x  x0 )  ( x  x0 )2  
dx x  x0 2 dx 2
xx 0
(1.53)
For functions that are relatively smooth, the magnitudes of the terms in this series
decrease as higher order derivatives are introduced, so an approximation of a
function can be achieved by selecting only the low-order terms. Choosing only
the first two, for example, we obtain
f ( x )  f ( x 0 )  f '( x 0 )( x  x 0 )
(1.54)
 f '( x 0 ) x  f ( x 0 )  f '( x 0 ) x 0
It can be seen from (1.54) that by keeping only the first two terms of the
Taylor series, an equation of a line results. The approximation of (1.54) will be
referred to as the linearization of f ( x ) . This linearization is illustrated in Figure
1.10. In the figure, the original curve f ( x ) appears as the wavy line. The point
of expansion, x  x 0 , is the point at which the curve is approximated as a straight
line by (1.54). For most curves, it is important that the straight-line approximation
not be used if the value of x strays too far from x 0 . However if x remains close
to x 0 , the Taylor series approximation is sometimes a very good one for practical
purposes. The accuracy and so-called “linear region” depend, of course, on the
particular function f ( x ) .
f ( x )  f ( x0 )  f ' ( x0 ) x  x0 
f ( x)
“linear region”
f ( x0 )
x
x0
Figure 1.10 Taylor series linearization of a curve.
When the function f depends on several variables, such as x1 , x2 ,, x n , not

only must all the partial derivatives of the individual variables x i be used in
(1.52), but all their cross-derivatives as well. That is, the Taylor series may be
expressed as
f ( x1 , x 2 ,  x n )  f ( x10 , x 20 , x n 0 )
FG
 ( x1  x10 )

   ( x n  x n0 )
 IJ
f ( x1 , x 2 ,, x n )
H x1 x n K xi  xi 0

1 FG
( x1  x10 )

   ( x n  x n0 )
 IJ 2
f ( x1 , x 2 , x n )
2! H x1 x n K xi  xi 0
 
Of course, for a linear approximation, only the first two terms of this series need
be retained.
It is also quite common that several functions of several variables need to be
linearized, such as
f1 ( x1 , x2 , x n )
f 2 ( x1 , x2 , x n )
(1.55)

f m ( x1 , x2 ,  x n )
In this case, each function f j ( x1 , x2 ,, x n ), j  1,, m can be expanded into a

Taylor series and thus linearized separately. Alternatively, we can use matrix-
vector notation and rewrite (1.55) as
LM f ( x , x , x ) OP
1 1 2 n
f ( x)  M
f ( x , x , x ) P
MM PP
2 1 2 n
(1.56)

N f ( x , x ,  x )Q
m 1 2 n
 
where f  f1 f2  fm and x  x1 x2  x n . Using this notation,
the linearized version of the nonlinear functions are (by taking the first two terms
of the Taylor series):
f ( x )
f ( x )  f ( x0 )  ( x  x0 ) (1.57)
x x0
bg
In this expression, the derivative of f x is a derivative of an m1 vector with
respect to an n 1 vector, resulting in an m  n matrix whose ( i, j )th element is
f i x j . See Appendix A.
1.2.2 Linearizing Differential Equations

The Taylor series linearization process can be performed on differential equations
as well as on functions. When each term of a nonlinear differential equation is
linearized in terms of the variables on which it depends, we say that the entire
equation has been linearized. When this has been done, all the linear analysis tools
presented here and elsewhere may be applied to the equation(s), remembering that
this linear equation is just an approximation of the original system. There are
some situations, such as in chaotic systems, wherein linear approximations are
not very good approximators of the true solution of the system, and still more,
such as when the nonlinear equations are not analytic (e.g., systems with static
friction), when the linearization does not apply at all. In many physical systems,
though, the two-term Taylor series approximation provides a reasonably accurate
representation of the system, usually good enough that linear controllers can be
applied, provided that the approximation is taken sufficiently near the operating
point.
Example 1.8: Linearization of a Differential Equation for an Inverted

Pendulum
The equation of motion can be derived for the model of an inverted pendulum on
a cart shown in Figure 1.11. In the model, ( t ) is the angle of the pendulum
clockwise with respect to the vertical, x ( t ) is the horizontal position of the cart
relative to some arbitrary fixed reference location, 2 is the length of the
pendulum, M and m are the masses of the cart and the pendulum, respectively,
and I is the moment of inertia of the pendulum about its center of gravity. F is a
force applied to the body of the cart.
2
( t )
mg
F
M
x (t )
Figure 1.11 Physical model of an inverted pendulum on a cart.
It can be shown that the two coupled differential equations that describe the
motion of this system are:
bm  M gx  m cos   m sin   F

2
(1.58)
e I  m j  mx cos   mg sin   0
2
Linearize these two equations using the first two terms of the Taylor series.
Solution:
Although only x ( t ) and ( t ) are considered “coordinate variables,” when
linearizing we consider their respective derivatives to also constitute independent
variables. Likewise, the input force F is a variable (although the equations already
happen to be linear in terms of F). We therefore are linearizing two equations,
each in terms of the seven independent variables ( t ),  ( t ),  ( t ), x ( t ), x ( t ), 
x( t )
, and F.
To begin, we select an operating point. For convenience, we choose the
“zero” position of  0 ( t )   0 ( t )  
 0 ( t )  0 , x0 ( t )  x 0 ( t )  
x0 ( t )  0 , and
F0  0 . We should note that these values are not all independent of one another.
If we arbitrarily select the first six values, then substitution into (1.58) allows us
to determine F0  0 . If, for example, the cart and pendulum were moving up a
wall, then a different F0 (nonzero) would have to be chosen in order to counteract
gravity. The equations in (1.58) must hold at all times.
We will apply (1.54) to the four terms in the first equation of (1.58) and leave
the details of the second equation as an exercise.
b
For m  M  x,g
bm  M gx  bm  M gx  bm  M gb x  x g
0 0
 0  b m  M g 
x
 b m  M g 
x
which of course returns the original expression since it was linear to begin with.
For m
 cos  ,

m
 cos   m
 0 cos  0 

m
 cos  d i b   g
0 0

0




dm
 cos  i d   i
0 0

0
 0  m bg
 0 sin  0   m cos  0 
 di
 m 

For m 2 sin  ,


m 2 sin   m 0 2 sin  0 

e
m 2 sin  j b   g
0 0
 0


e
  m 2 sin  j d   i
0 0
 0
bg
 0  m 0 2 cos  0   2m 0 sin  0  di
0
It may at first seem somewhat surprising that this term should disappear entirely.
However, recalling that the equilibrium point we have chosen has  0  0 , the
“linear” region in this problem will include only small velocities. Because the
velocity in this term appears only as the square, and we are assuming that second
order and higher terms are negligible, this entire term should be negligible near
the equilibrium.
As for the final term, F, it is clear that this is already linear, and we will leave
it as is. The “linearized” version of the first of Equations (1.58) is therefore:
bm  M gx  m  F (1.59)
Note that this equation has two second-order derivatives in it, which are the
highest derivatives of each variable ( t ) and x ( t ) . The second equation will
linearize to
( I  m 2 )
  m
x  mg  0 (1.60)
It also has two second-order derivatives. In order to construct a state space

representation for such a system, in both Equation (1.59) and (1.60) we will have
solve for one such second derivative and substitute it into the other, so that each
equation contains only one of the highest derivatives of any of the state variables.
This procedure and the subsequent formulation of the state equations is left as
Exercise 1.12.
The Choice of Operating Point

One should be careful in applying these methods with the goal of arriving at an
equation of the form
f ( x )
x  f ( x0 )  ( x  x0 ) (1.61)
x x0
If in (1.61) we have
f ( x )
f ( x0 )  ( x0 )  0
x x0
then (1.61) is not linear! It has a constant offset and will therefore violate the
homogeneity and additivity conditions required for linearity. Equation (1.61) will
indeed look like a straight line on the x versus x plane, but it will not pass
through the origin of this plane. Such a system is called affine.
In order to make the system linear, one must choose an appropriate operating
value for the variable x . In particular, choose as the operating point the
equilibrium value x e such that f ( x e )  0 . Then to eliminate the constant term

in (1.61), we must make a change of variable   x  xe . Then, x    x e and
x   . Substituting these values into (1.61), we get
f ( x ) f ( x )
  f ( xe )  ()   (1.62)
x xe x xe
This equation is then truly linear. If after solving for  , we desire to know the
value of x , then the change of variables may be reversed.
1.3 Summary
In this chapter we have introduced the state space notation to simple models of
physical systems. Our purpose has been to establish terms and definitions that are
both common to and different from the transfer function descriptions with which
the reader may be more familiar. Using the state variable technique, our main
points have been as follows:
• State variables are simply regarded as a set of variables in which the
behavior of a system may be mathematically modeled.
• Mathematical models for use in state variable systems may be derived,
simplified, linearized, and discretized just as any other mathematical
model is produced. It is only the choice of state variables that is new so
far.
• State variables and state variable equations together constitute the same
information about a physical system as do the original differential
equations derived from familiar conservation and constitutive laws.
• State variables may be changed to represent either other physical

variables or variables that have no discernible physical significance at
all. This changes the structure of the matrix elements in the state
equations but does not alter the fact that the equations describe the same
physical system equally well.
We have tried to stress the need for matrix and vector analysis and linear
algebra in the discussions of this chapter. In the next chapter, we will begin to
study these mathematical tools. They will later lead us to see the benefits and
drawbacks of the state variable techniques.
1.4 Problems
1.1 A system has an input u( t ) and an output y ( t ) , which are related by the
information provided below. Classify each system as linear or nonlinear
and time-invariant or time-varying.
a) y ( t )  0 for all t .
b) y ( t )  a , a  0 , for all t .
c) y ( t )  3u( t )  2 .
d)
y(t )
a
u (t )
a
e)
y (t )
u (t )
f) y ( t )  3e  t y ( t )  y ( t )  u( t ) .
g) y ( t )  (1  y 2 ( t )) y ( t )   2 y ( t )  u( t ) ,   constant  0 .
h) y ( t )  u( t  3) .
i) y ( t )  u( t ) y ( t )  0 .
z
t
j) y( t )  e   u( t   )d .
0
k) y ( k  2 )  0.4 y ( k  1)  y ( k )  3u( k  2 )  u( k ) .
k
l) y( k )   sin( iT ) e  iT u( k  i ) , T  constant  0 .
i0
1.2 Figure P1.2 shows a model commonly used for automobile suspension
analysis. In it, the uneven ground specifies the position of the wheel’s
contact point. The wheel itself is not shown, as its mass is considered
negligible compared to the mass of the rest of the car. Write a differential
equation and a state variable description for this system, considering the
height of the car, x ( t ) , to be the output, and the road height, y ( t ) , to be
the input.
M x (t )
B K
y (t )
y0
P1.2
1.3 The motor in Figure P1.3 exerts a torque  on the shaft of a motor, which
has inertia J m and bearing friction Bm (assumed viscous). This motor is
attached to a load inertia J L , which itself has a viscous friction BL . The
motor coupling is slightly flexible and is modeled as a torsional spring K
. Write the equations of motion and a set of state equations for the system,
taking ( t ) as the input, and  L ( t ) as the output.
( t )  m (t )  L (t )
K
Jm JL
Bm BL
P1.3
1.4 For the mechanical system in Figure P1.4 with a spring, friction forces, and
an external force, write state equations using the following sets of
variables:
a) State variables 1  x1 ,  2  x1 ,  3  x 2 , and  4  x 2 ; and output
variable y  x 2 .
b) State variables 1  x1 ,  2  x1 ,  3  x 2  x1 , and  4  x 2  x1 ;
and output variable y  x 2  x1 .
x2 x1
K
F
m2 B2
B1
m1
P1.4
1.5 For the circuit shown in Figure P1.5, choose appropriate state variables and
write state equations, taking as output the voltage across the resistor.
L C
R v R (t )
v (t )
P1.5
1.6 For the circuit shown in Figure P1.6, write a single second-order
differential equation in terms of the voltage v x ( t ) and the input x ( t ) . Then
write state equations using state variables v C ( t ) and iL ( t ) , where the
output of the system is considered to be v x ( t ) , and the input is x ( t ) .
C
i L (t )
v C (t ) R
x (t ) L v x (t )
P1.6
1.7 For the circuit shown in Figure P1.7, with ideal OP-AMPS, find a state
variable representation, using the capacitor voltages as state variables, and
the signal x ( t ) as the output. What will be the zero-input solution for x ( t )
?
R
R 10F
10F
u( t ) 10k
10k
x (t )
R
P1.7
1.8 Write two differential equations that describe the behavior of the circuit of
Example 1.2, using v L ( t ) and v C ( t ) as the state variables.
1.9 Given the state variable description of the system in terms of x i ( t ) below,
change the state variables and write new state equations for the variables
1 ( t )  3x1 ( t )  2 x2 ( t ) , and  2 ( t )  7 x1 ( t )  5x 2 ( t ) .
LM x (t )OP  LM 11
1 OP LM x (t ) OP  LM0OPu(t )
10 1
Nx (t )Q N21
2 18Q N x ( t )Q N1Q
2
y(t )  2
L x (t )OP  u(t )
2 M
1
Nx (t )Q
2
1.10 For the state variable description of the system in terms of x i ( t ) ,
LM x (t )OP LM 18
1 9 OP LM x (t ) OP LM1OP
13 1
MMx (t )PP  MM 50 23 35 P M x ( t )P  M 0 Pu( t )

31 46PQ MN x ( t )PQ MN 1 PQ
2 2
N x (t )Q N65
3 3
LM x (t )OP1
y( t )  5 5 5 M x ( t )P
MN x (t )PQ
2
3
change the state variables and write new state equations for variables
1 ( t )  4 x1 ( t )  2 x2 ( t )  3x3 ( t )
 2 ( t )  15 x1 ( t )  7 x2 ( t )  10 x3 ( t )
and
 3 ( t )  5 x1 ( t )  2 x2 ( t )  3x3 ( t )
1.11 The robot shown in Figure P1.11 has the differential equations of motion
given. Symbols m1 , m2 , I1 , I 2 , 1 , and g are constant parameters,
representing the characteristics of the rigid body links. Quantities
1 and d 2 are the coordinate variables and are functions of time. The inputs
are 1 and  2 . Linearize the two equations about the operating point
    
1 1  0, d  3 , and d  d  0 .
1 2 2 2
d2
1 m2
m1 1
P1.11
1  2m2 d2 1d2  ( m11  m2 d2 ) g cos 1  1

( m112  I1  I2  m2 d22 )
m d  m d  2  m g sin   
2 2 2 2 1 2 1 2
1.12 For the system in Example 1.8, the nonlinear equations of motion are given
in (1.58). Show that (1.60) is the linear approximation of the second of
these equations, and combine the two linearized equations (1.59) and
(1.60) into state space form.
1.13 A permanent-magnet DC motor with a connected inertial and friction load

is depicted in Figure P1.13. The motor armature is driven with voltage V,
and the motor turns through angle  with torque  . The armature has
resistance R and inductance L, and the armature current is denoted i. The
mechanical load is an inertia J, and the bearings have viscous friction
coefficient f. The motor produces a back-emf of e  kb and a torque of
  ka i , where k a and k b are constants. Determine a set of describing
equations for the electrical and mechanical components of the system.

Then express these equations as a set of linear state equations, using the
voltage V as input and the angle  as output.
R L
 
V (t )
e(t ) J f
i (t )
P1.13
1.5 References and Further Reading

The formulation of the dynamic equations for physical systems can be found in
any physics text. For information specific to mechanical systems, texts in statics
and dynamics can be consulted, and for circuit analysis, there are a great many
elementary texts, including [1], which focuses on state variable analysis of
electrical networks. A reference that ties them together and introduces the unified
modeling terminology of “through” variables and “across” variables is [4]. Other
good introductions to the state space representation for physical systems can be
found in [2], [7], [8], and [10]. In particular, [10] gives a very detailed introduction
to linear system terminology and definitions. For systems described in frequency-
domain, which we do not treat in much depth in this book, the student can consult
[3] and [7].
Additional state variable models can be found in [5] and [8], both of which
provide numerous examples from systems that engineering students do not
traditionally encounter, such as genetics, populations, economics, arms races, air
pollution, and predator-prey systems.
Further information on nonlinear systems and linearization is given in [9].
[1] Belevitch, V., Classical Network Theory, Holden-Day, 1968.
[2] Brogan, William L., Modern Control Theory, 3rd edition, Prentice-Hall, 1991.
[3] Callier, Frank M., and Charles A. Desoer, Multivariable Feedback Systems,
Springer-Verlag, 1982.
[4] Cannon, Robert H. Jr., Dynamics of Physical Systems, McGraw-Hill, 1967.
[5] Casti, John L., Linear Dynamical Systems, Academic Press, 1987.
[6] Franklin, Gene, and J. David Powell, Digital Control of Dynamic Systems,
Addison-Wesley, 1981.
[7] Kailath, Thomas, Linear Systems, Prentice-Hall, 1980.
[8] Luenberger, David G., Introduction to Dynamic Systems, John Wiley & Sons,
1979.
[9] Slotine, Jean-Jacques, and Weiping Li, Applied Nonlinear Control, Prentice-Hall,
1991.
[10] Zadeh, Lotfi A. and Charles A. Desoer, Linear System Theory: The State Space
Approach, McGraw-Hill, 1963.

Models of Linear Systems

Uploaded by

Copyright:

Available Formats

Models of Linear Systems

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Models of Linear Systems

Uploaded by

Copyright:

Available Formats

1

Models of Linear Systems

1.1 Linear Systems and State Equations

Figure 1.1 Elementary representation of a system acting on an input and producing

1.1.1 Definitions and Review

Memory: A system with memory is one whose output depends

Systems with memory most often occur as differential equations (continuous-

Causality: A system is said to be causal if the value of the

for all t up to time t 0 but no further, i.e., only for t  t 0

Time Invariance: Given an initial time t 0 , the output of a

Time-varying systems are typically systems in which time appears as an

Linearity: A linear system is one that satisfies homogeneity

1.1.2 Physical System Modeling

F all flow variablesIJ   FG net equivalent flowIJ

zero, because it is impossible for net charges to accumulate indefinitely in a wire

F signed changes in a potential I

At any such component, these relationships can be used to substitute a flow

displaced relative to a fixed reference defined elsewhere. The constitutive

Example 1.1: Mechanical System Equations

Applying Newton’s law, ma   F , we get

Rearranging these equations to a more convenient form,

Example 1.2: Electrical System Equations

Figure 1.6 Electric circuit example with variables v c ( t ) and i( t ) .

1.1.3 State Equations

We will find it convenient to express such a system of equations in vector-

If in addition, one of the state variables, say 1 ( t ) , is designated as the “output”

Example 1.3: State Variables for the Mechanical System Example

Example 1.4: State Variables for the Electrical System Example

LM OP  LM 1CR 1 OP L  O L 0 O

Alternative State Variables

we would have required derivatives of u( t ) in the state equations. According to

x ( t )  a1 x ( t )  a0 x ( t )  b2 u( t )  b1u( t )  b0 u( t ) (1.23)

Figure 1.8 A second type of simulation diagram and state equations.

Changing State Variables

m1x1  Bx2  Kx2  F

variables, we can solve the first equation of (1.27) for 

Here again we have let the output variable be y ( t )  x 2  x1 (   x 2 ( t )    3 ) .

we can use vector-matrix notation immediately to write

LMv R OP LM0 R2OP LMv OP c

Alternatively, using the inverse relationship,

LMv OP  LM0 R2OP LMv OP  LM R1  0R OP LMv OP  LM 11

LM0 R OP LM CR1 OP L 1

1.1.4 Discrete-Time Systems

where k is an integer. In such a situation, we will sometimes simplify the notation

where the notation Ad and bd are introduced to distinguish these discrete-time

Example 1.5: Direct Difference Equation Modeling of a Savings Account

This is a first-order difference equation and, hence, is already in linear state

Example 1.6: A Difference Equation for a Predator-Prey System

where k1 and k 2 are constants of proportionality. Of course, Pp ( 0 ) and Pg ( 0 )

LM P ( d  1)OP L1  k k1OP LM P ( d )OP LM0OP

Example 1.7: Discretized Differential Equations for a Harmonic Oscillator

Substituting these approximations into the original Equation (1.43), we get

or simplifying and dropping the T ’s in the time arguments,

This can be seen as a discrete-time equation of the form (1.39).

MNx ( k  2)PQ  MN( T

1.1.5 Relationship to Transfer Functions

If we assume that the initial condition is x ( t 0 )  0 for this system, Equations

Simplifying the first of these equations,