Lecture 13: State Space Representation,

Discrete-time State-Space Models,

Stability,
Lyapunov Theory

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
Lyapunov Stability

25
26
⚫ We do not propose P and hope to find Q.

⚫ We propose Q and check if there exist a postive

definite P that satisfies 𝐴𝑇 𝑃 + 𝑃𝐴 < 0, namely
such that 𝐴𝑇 𝑃 + 𝑃𝐴 is negative definite.

⚫ If such a P exist, the system is stable.

27
Lecture 14_2: Controlability,
Observability, Pole-Placement

28
Conrolability and Observability

• Controllability is concerned with the problem of whether it is possible to steer a

system from a given initial state to an arbitrary state:

➢ a system is said to be controllable if it is possible by means of an unbounded

control vector to transfer the system from any initial state to any other state in a
finite number of sampling periods.

• Observability is concerned with the problem of determining the state of a

dynamic system from observations of the output and control vectors in a finite
number of sampling periods.

➢ A system is said to be observable if, with the system in state x(O), it is possible
to determine this state from the observation of the output and control vectors
over a finite number of sampling periods.

The concepts of controllability and observability were introduced by R. E. Kalman.

29
➢ Note that the concept of controllability is the basis for the solutions of the pole
placement problem and the concept of observability plays an important role for the
design of state observers.
➢ The design approach of placing the closed-loop poles in the desired locations in the z
plane is called the pole placement design technique; that is, in the pole placement
design technique we feed back all state variables so that all poles of the closed-loop
system are placed at desired locations.
➢ ln practical control systems, however, measurement of all state variables may not be
possible; in that case, not all state variables will be available for feedback. To
implement a design based on state feedback, it becomes necessary to estimate the
unmeasurable state variables. Such estimation can be done by use of state observers.
➢ The pole placement design process of control systems may be separated into two
phases. In the first phase, we design the system assuming that all state variables are
available for feedback. In the second phase, we design the state observer that estimates
all state variables (or only those that are not directly measurable) that are required for
feedback to complete the design

30
 Controllability
➢ A control system is controllable if every state variable can be controlled in a finite time
period by some unconstrained control signal, If any state variable is independent of the
control signal, then it is impossible to control this state variable and therefore the system
is uncontrollable.
Complete State Controllability for a Linear Time-Invariant Discrete Time Control System.
Consider the discrete-time control system defined by
𝒙((𝑘 + 1)𝑇) = 𝑮𝒙(𝑘𝑇) + 𝑯𝒖(𝑘𝑇)
where
𝒙(𝑘𝑇) = state vector (n-vector) at kth sampling instant
𝒖(𝑘𝑇) = control signal at kth sampling instant
𝑮 = n x n matrix
H = n x 1 matrix
T = sampling period
We assume that 𝒖(𝑘𝑇) is constant for 𝑘𝑇 < 𝑡 < (𝑘 + 𝑙)𝑇. Necessary condition for
complete state controllability;

𝒓𝒂𝒏𝒌 𝑯 𝑮𝑯 … 𝑮𝒏−𝟏 𝑯 = 𝒏 (can span the n-dimensional space)

[𝑯 𝑮𝑯 … 𝑮𝒏−𝟏 𝑯] is called Controlability Matrix.

31
Summary: It is said to be state controllable if there exists a piecewise constant control
signal 𝑢(𝑘𝑇) defined over a finite number of sampling periods such that, starting from any
initial state, the state 𝑥(𝑘𝑇) can be transferred to the desired state 𝑥𝑓 in at most n sampling
periods.
Alternative Form of the Condition for Complete State Controllability
If the eigenvectors of G are distinct, then it is possible to find a transformation matrix P
such that

Note that if the eigenvalues of G are distinct then the eigenvectors of G are distinct.
However, the converse is not true. (For example, an n x n real symmetric matrix having
multiple eigenvalues has n distinct eigenvectors.)

➢ Hence, the condition for complete state controllability is that, if the eigenvectors of G
are distinct, then the system is completely state controllable if and only if no row of
𝑷−𝟏 𝑯 has all zero elements.

32
 If the G matrix does not possess distinct eigenvectors, then diagonalization is
impossible. In such a case, we may transform G into a Jordan canonical form.

The 3x3 and 2x2 submatrices on the main diagonal are called Jordan blocks. Suppose it
is possible to find a transformation matrix S such that 𝑺−1 𝐺𝑺 = 𝑱.

➢ The system is completely state controllable if and only if (1) there isn’t any two
Jordan blocks in J are associated with the same eigenvalues, (2) the elements of any
row of 𝑺−1 H that corresponds to the last row of each Jordan block are not all zero,
and (3) the elements of each row of 𝑺−1H that correspond to distinct eigenvalues are
not all zero.

33
34
Condition for Complete State Controllability in the z Plane. The condition for complete
state controllability can be stated in terms of pulse transfer functions, A necessary and
sufficient condition for complete state controllability is that no cancellation occur in the
pulse transfer function. For example following system is not state controlable
𝑌(𝑧) 𝑧 + 0.2
=
𝑈(𝑧) (𝑧 + 0.8)(𝑧 + 0.2)
The same conclusion can be obtained, of course, by writing this pulse transfer
function in the form of state equations. A possible state-space representation for this
system is

(rank([𝐻 𝐺𝐻]) =1)

35
Complete Output Controllability. In the practical design of a control systems, we may
want to control the output rather than the state of the system. Complete state
controllability is neither necessary nor sufficient for controlling the output of the
system. For this reason, it is necessary to define separately complete output
controllability.
Consider the system defined by the equations

𝒙((𝒌 + 1)𝑇) = 𝑮𝒙(𝑘𝑇) + 𝑯𝒖(𝑘𝑇)

𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)
where
x(kT) = state vector (n-vector) at kth sampling instant
u(kT) = control signal (scalar) at kth sampling instant
y(kT) = output vector (m-vector) at kth sampling instant
G = n x n matrix
H = n x 1 matrix
C = m x n matrix

36
Thus, as in the case of complete state controllability, a necessary and sufficient
condition for the system to be completely output controllable is that colum vectors
of [𝑪𝑯, 𝑪𝑮𝑯, … , 𝑪𝑮𝒏−𝟏 𝑯] span the m-dimensional output space, or that

𝒓𝒂𝒏𝒌[𝑪𝑯, 𝑪𝑮𝑯, … , 𝑪𝑮𝒏−𝟏 𝑯] = 𝒎

• Complete state controllability implies complete output controllability if and only if the
m rows of C are linearly independent.
Consider the system defined by the equations
𝒙 𝒌 + 1 𝑇 = 𝑮𝒙 𝑘𝑇 + 𝑯𝒖 𝑘𝑇
𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)+ D𝑢(𝑘𝑇)
where
x(kT) = state vector (n-vector) at kth sampling instant
u(kT) = control signal (scalar) at kth sampling instant
y(kT) = output vector (m-vector) at kth sampling instant
G = n x n matrix
H = n x 1 matrix
C = m x n matrix
D = m x r matrix

37
 Thus, as in the case of complete state controllability, a necessary and sufficient
condition for the system to be completely output controllable is that colum vectors of
[𝑫 𝑪𝑯, 𝑪𝑮𝑯, … , 𝑪𝑮𝒏−𝟏 𝑯] span the m-dimensional output space, or that

𝒓𝒂𝒏𝒌[𝑫 𝑪𝑯, 𝑪𝑮𝑯, … , 𝑪𝑮𝒏−𝟏 𝑯] = 𝒎

It is noted that the presence of matrix D in the system output equation always helps to
establish complete output controllability.

Observability
Consider the system defined by the equations
𝒙 𝒌 + 1 𝑇 = 𝑮𝒙 𝑘𝑇
𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)
where
x(kT) = state vector (n-vector) at kth sampling instant
y(kT) = output vector (m-vector) at kth sampling instant
G = n x n matrix
C = m x n matrix

38
➢ The system is said to be completely observable if every initial state 𝑥(0) can be
determined from the observation of 𝑦(𝑘𝑇) over a finite number of sampling periods.

➢ The system, therefore, is completely observable if every transition of the state eventually
affects every element of the output vector.

➢ The concept of observability is useful in solving the problem of reconstructing

unmeasurable state variables.

➢ It will be seen later that state feedback control systems designed by the pole placement
method will require feedback of weighted state variables.

➢ In practice, however, the difficulty encountered with state feedback control systems is
that some of the state variables are not accessible for direct measurement. Then it
becomes necessary to estimate the unmeasurable state variables in order to construct the
feedback control signals.

39
The reason we are considering the unforced system is as follows. If the system is described
by the equations
𝒙((𝑘 + 1)𝑇) = 𝑮𝒙(𝑘𝑇) + 𝑯𝒖(𝑘𝑇)
𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)+D𝒖(𝑘𝑇)

Then
𝒙 𝑘𝑇 = 𝑮𝒌 𝒙 0 + σ𝒌−𝟏
𝒋=𝟎 𝑮
𝒌−𝒋−𝟏 𝑯𝒖 𝒋𝑻

and
𝒚 𝑘𝑇 = 𝑪𝑮𝒌 𝒙 0 + σ𝒌−𝟏
𝒋=𝟎 𝑪𝑮
𝒌−𝒋−𝟏 𝑯𝒖 𝒋𝑻 + 𝑫𝒖(𝒌𝑻)

Since the matrices G, H, C, and D are known and 𝑢(𝑘𝑇) is also known, the second and
third terms on the right-hand side of this last equation are known quantities. Therefore, they
may be subtracted from the observed value of 𝑦(𝑘𝑇). Hence, for investigating a necessary
and sufficient condition for complete observability, it suffices to consider the system
described by Equations,
𝒙 𝒌 + 1 𝑇 = 𝑮𝒙 𝑘𝑇
𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)

40
Complete Observability of Discrete Time Systems.
The system is completely observable if, given the output 𝒚(𝑘𝑇) over a finite number of
sampling periods, it is possible to determine the initial state vector 𝒙(0).

Since the solution x(kT) ,

𝒙(𝑘𝑇) = 𝑮𝒌 𝒙(0)
then
𝒚 𝑘𝑇 = 𝑪𝑮𝒌 𝒙 0

Complete observability means that, given y(0), y(T), y(2T), ..., it is possible to determine
𝒙1 0 , 𝒙𝟐 (0), . . . , 𝒙𝒏 (0). To determine n unknowns, we need only n values of y(kT).
Hence, we may use the first n values of 𝒚(𝑘𝑇), or 𝒚(0), 𝒚(𝑇), . . . , 𝒚((𝑛 − 1)𝑇) for the
determination of 𝒙1 0 , 𝒙2 (0), . . . , 𝒙𝑛 (0).

41
For a completely observable system, given
𝒚(0) = 𝑪𝒙(𝟎)
𝒚(𝑇) = 𝑪𝑮𝒙(𝟎)

𝒚((𝑛 − 𝑙)𝑇) = 𝑪𝑮𝒏−𝟏 𝒙(𝟎)

we must be able to determine 𝒙𝟏 0 , 𝒙2 (0), . . . , 𝒙𝑛 (0). Noting that y(kT) is an m vector, the
preceding n simultaneous equations yield nm equations, all involving
𝒙𝟏 0 , 𝒙2 (0), . . . , 𝒙𝑛 (0). To obtain a unique set of solutions 𝒙1 0 , 𝒙2 (0), . . . , 𝒙𝒏 (0)) from
these nm equations, we must be able to write exactly n linearly independent equations among
them. This requires that the nm x n matrix

be of rank n. Noting that the rank of a matrix and that of the conjugate transpose of the
matrix are the same, it is possible to state the condition for complete observability as follows
𝑟𝑎𝑛𝑘 𝑪∗ 𝑮∗ 𝑪∗ … (𝑮∗ 𝑛−1 𝑪∗ ]
=𝑛
𝑪∗ 𝑮∗ 𝑪∗ … (𝑮∗ 𝒏−𝟏 𝑪∗ ] is
called the observability matrix.

42
Alternative Form of the Condition for Complete Observability: Suppose the eigenvalues of G
are distinct, and a transformation matrix P transforms G into a diagonal matrix, so that
P −𝟏 𝑮𝑷 is a diagonal matrix. Let us define 𝒙 𝑘𝑇 = 𝑷𝒙 𝑘𝑇 ,
therefore
𝒙 𝑘 + 1 𝑇 = 𝑷−𝟏 𝑮𝑷𝒙 𝑘𝑇
𝒚 𝑘𝑇 = 𝑪𝑷𝒙 𝑘𝑇
Thus
𝒏
𝒚(𝑛𝑇) = 𝑪𝑷 𝑷−𝟏 𝑮𝑷 𝒙(0)

Which is as follows in extended form,

The system is completely observable if and only if none of the columns of the m x n matrix
CP consists of all zero elements. This is because, if the 𝒊𝑡ℎ column of CP consists of all zero
elements, then the state variable 𝒙(𝟎) will not appear in the output equation and therefore
cannot be determined from observation of 𝒚(𝑘𝑇).

43
If the matrix G involves multiple eigenvalues and cannot be transformed into adiagonal
matrix, then by using a suitable transformation matrix S we may transform G into the
Jordan canonical form:
𝑺−𝟏 𝑮𝑺 = 𝑱
where J is in the Jordan canonical form. Let us define x(kT) = S x(kT)

𝒙 𝑘 + 1 𝑇 = 𝑺−𝟏 𝑮𝑺𝒙 𝑘𝑇 = 𝑱 x(kT)

𝒚 𝑘𝑇 = 𝑪𝑺 𝑘𝑇
Thus
𝒏
𝒚(𝑛𝑇) = 𝑪𝑺 𝑺−𝟏 𝑮𝑺 𝒙(0)

The system is completely observable if and only if (1) no two Jordan blocks in J are
associated with the same eigenvalue, (2) none of the columns of CS that corresponds to
the first row of each Jordan block consists of all zero elements, and (3) no columns of
CS that correspond to distinct eigenvalues consist of all zero elements.

44
45
46
Condition for Complete Observability in the z Plane. The condition for complete
observability can also be stated in terms of pulse transfer functions, A necessary and
sufficient condition for complete observability is that no pole-zero cancellation occur in
the pulse transfer function. If cancellation occurs, the canceled mode cannot be observed
in the output.
For example following system is not completely observable
𝑌(𝑧) (𝑧 + 1)(𝑧 + 4)
=
𝑈(𝑧) (𝑧 + 1)(𝑧 + 2)(𝑧 + 3)
The same conclusion can be obtained by following analysis. A state-space represenatation of
the system is 𝒙((𝒌 + 1)𝑇) = 𝑮𝒙(𝑘𝑇) + 𝑯𝒖(𝑘𝑇)
𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)

𝒓𝒂𝒏𝒌 𝑪∗ 𝑮∗ 𝑪∗ … (𝑮∗ 𝒏−𝟏 𝑪∗ ]

with is less then 3.

47
 Duallity
Principle of Duality. In what follows, we examine the relationship between controllability
and observability. Consider the system S1 defined by the equations
𝒙((𝒌 + 1)𝑇) = 𝑮𝒙(𝑘𝑇) + 𝑯𝒖(𝑘𝑇)
𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)
where
x(kT) = state vector (n-vector) at kth sampling instant
u(kT) = control signal (r-vector) at kth sampling instant
y(kT) = output vector (m-vector) at kth sampling instant
G = n x n matrix
H = n x r matrix Dual counterpart of S1, which we call system S2, defined by
C = m x n matrix the equations
ෝ ((𝒌 + 1)𝑇) = 𝑮∗ 𝒙
𝒙 ෝ(𝑘𝑇) + 𝑪∗ 𝒖(𝑘𝑇)
𝒚(𝑘𝑇) = 𝑯∗ 𝒙 ෝ(𝑘𝑇)
where
ෝ(kT) = state vector (n-vector) at kth sampling instant
𝒙
u(kT) = control signal (m-vector) at kth sampling instant
y(kT) = output vector (r-vector) at kth sampling instant
G = n x n matrix
H = n x m matrix
C = r x n matrix
48
We shall now examine an analogy between controllability and observability of S1 and S2.
This analogy is referred to as the principle of duality, due to Kalman.

➢ The principle of duality states that system S1 defined is completely state controllable
(observable) if and only if system S2 defined completely observable (state controllable).

To verify this principle, let us write down the necessary and sufficient conditions for
complete state controllability and complete observability for systems S1 and S2,
respectively.

FOR SYSTEM 𝑺𝟏
A necessary and sufficient condition for complete state controllability is that
𝑟𝑎𝑛𝑘 [𝑯: 𝑮𝑯: · · ·∶ 𝑮𝑛−1 𝑯] = 𝑛
A necessary and sufficient condition for complete observability is that
𝑟𝑎𝑛𝑘 𝑪∗ 𝑮∗ 𝑪∗ … (𝑮∗ 𝑛−1 𝑪∗ ] =𝑛

49
FOR SYSTEM 𝑺𝟐
necessary and sufficient condition for complete state controllability is that

𝑟𝑎𝑛𝑘 𝑪∗ 𝑮∗ 𝑪∗ … (𝑮∗ 𝑛−1 𝑪∗ ] =𝑛

A necessary and sufficient condition for complete observability is that

𝑟𝑎𝑛𝑘 [𝑯 𝑮𝑯: · · · 𝑮𝑛−1 𝑯] = 𝑛

By comparing these conditions, the truth of the principle of duality is apparently.

➢ We see that system S1 being completely state controllable is equivalent to system S2

being completely observable.
➢ And system S1 being completely observable is equivalent to system S2 being completely
state controllable.

By use of this principle, the observability of a given system can be checked by testing the
state controllability of its dual.

50
 DESIGN VIA POLE PLACEMENT

➢ We assume that all state variables are measurable and are available for feedback.
With this assumption, if the system considered is completely state controllable, then
poles of the closed-loop system may be placed at any desired locations by means of
state feedback through an appropriate state feedback gain matrix.
➢ The technique begins with a determination of the desired closed-loop poles based on
transient-response and/or frequency-response requirements such as speed, damping
ratio, or bandwidth. Given such considerations, let us assume that we decide that the
desired closed-loop poles are determined.
➢ A proper sampling period should be selected. In choosing the sampling period, care
must be exercised so that the desired system will not require unusually large control
signals. Otherwise, saturation phenomena will occur in the system. If saturation takes
place in the system, the system will become nonlinear, and the design method will no
longer apply.
➢ Then, by choosing an appropriate gain matrix for state feedback, it is possible to
make the system to have closed-loop poles at the desired locations, provided that the
original system is completely state controllable.

51
(a) Open-loop control system

(b) closed-loop control system with 𝒖(𝑘) = −𝑲𝒙(𝑘)

52
Consider the system defined by the equations

𝒙((𝒌 + 1)𝑇) = 𝑮𝒙(𝑘𝑇) + 𝑯𝒖(𝑘𝑇)

where
x(kT) = state vector (n-vector) at kth sampling instant
u(kT) = control signal (scalar) at kth sampling instant
G = n x n matrix
H = n x 1 matrix
We assume that the magnitude of the control signal u(k) is unbounded. If the control
signal u(k) is chosen as
𝑢(𝑘) = −𝑲𝒙(𝑘)
where K is the state feedback gain matrix (a 1 x n matrix), then the system becomes a
closed-loop control system and its state equation becomes
𝒙(𝑘 + 1) = (𝑮 − 𝑯𝑲)𝒙(𝒌)
Note that we choose matrix K such that the eigenvalues of G - HK are the desired
closed-loop poles. Such a K matrix exists if and only if the system is the system be
completely state controllable. That is to say, the necessary and sufficient condition for
arbitrary pole placement is that the system be completely state controllable.

53
Ackermann's Formula. Let assume that the system is completely state controllable.
By using the state feedback 𝑢(𝑘) = −𝑲𝒙(𝒌), we wish to place closed-loop poles at
𝑧 = µ1 , 𝑧 = µ2 , . . , 𝑧 = µ𝑛 . That is, we desire the characteristic equation to be

෩ = 𝑮 − 𝑯𝑲. Since the Cayley-Hamilton theorem states that G satisfies its

Let us define 𝑮
own characteristic equation, we have

Cayley-Hamilton theorem: A square matrix is satisfies its own characteristic polynomial

equation.
I=I
෩ = 𝑮 − 𝑯𝑲
𝑮
෩ 𝟐 = 𝑮 − 𝑯𝑲 𝟐 = 𝐆𝟐 − 𝐆𝐇𝐊 − 𝐇𝐊𝑮෩
𝑮
෩ 𝟑 = 𝑮 − 𝑯𝑲 𝟑 = 𝐆𝟑 − 𝐆𝟐 𝐇𝐊 − 𝐆𝐇𝐊𝑮෩ − 𝐇𝐊𝑮
𝑮 ෩𝟐
⋮
෩ = 𝑮 − 𝑯𝑲 = 𝐆 − 𝐆𝒏−𝟏 𝐇𝐊 − 𝐆𝐇𝐊𝑮෩ − 𝐇𝐊𝑮
𝑮 𝒏 𝒏 𝒏 ෩ 𝒏−𝟏

54
Multiplying the above equations in order by 𝜶𝒏 , 𝜶𝒏−𝟏 , … , 𝜶𝟏 , 𝜶𝟎 , respectively, and adding
the results, we obtain

which can be written as follows:

Notice that

55
Therefore,

Since the system is completely state controllable, the controllability matrix

[𝑯: 𝑮𝑯: · · ·∶ 𝑮𝑛−1 𝑯]
is of rank n and its inverse exists. Therfore we can write the following,

56
Premultiplying both sides of this last equation by [0 0 0 … 0 1] yields

which can be simplified to

This expression for matrix K is called Ackermann's formula.

57
Example: Consider the system
𝒙(𝒌 + 1) = 𝑮𝒙(𝒌) + 𝑯𝒖(𝒌)
where

Note that

Hence

Determine a suitable state feedback gain matrix K such that the system will have the
closed-loop poles at

Let us first examine the rank of the controllability matrix

The rank of it is 2.

58
Thus, the system is completely state controllable, and therefore arbitrary pole placement is
possible. The characteristic equation for the desired system is

𝒛𝑰 − 𝑮 + 𝑯𝑲 = 𝑧 − 0.5 − 𝑗05 𝑧 − 05 + 𝑗0.5 = 𝑧 2 − 𝑧 + 0.5 = 0

𝛼1 = −1 𝛼2 = 0.5
Referring to Ackermann's formula given
𝑲 = 𝟎 𝟏 𝑯 𝑮𝑯 −𝟏 𝝓(𝑮)

where

Thus,

59
 Observer Design

A state observer, also called a state estimator, is a subsystem in the control system that
performs an estimation of the state variables based on the measurements of the output and
control variables,

➢ Full-order state observation means that we observe (estimate) all n state variables
regardless of whether some state variables are available for direct measurement.

➢ There are times when this will be unnecessary, when we will need observation of only
the unmeasurable state variables but not of those that are directly measurable.
Observation of only the unmeasurable state variables is referred to as minimum-order
state observation,

➢ Observation of all unmeasurable state variables plus some (but not all) of the measurable
state variables is referred to as reduced-order state observation.

➢ The necessary and sufficient condition for state observation is that the system be
completely observable.

60
Thus, for a completely observable system, the state vector can be determined in at most n
sampling periods. In the presence of external disturbances and measurement noises,
however, this approach may not give an accurate determination of the state vector. Hence,
to determine the state vector in the presence of disturbances and measurement noises, a
different approach is necessary.

Let us assume that the state 𝑥(𝑘) is to be approximated by the state 𝑥(𝑘)
෤ of the dynamic
model:
𝑥෤ 𝑘 + 1 = 𝐺 𝑥෤ (𝑘) + 𝐻𝑢(𝑘)
𝑦(𝑘) = 𝐶 𝑥෤ (𝑘)
where matrices G, H, and C are the same as those of the original system, Also, let us
assume that the dynamic model is subjected to the same control signal 𝑢(𝑘) as the original
system. If the initial conditions for the actual system and the dynamic model defined are
the same, then the state 𝑥(𝑘) and the state 𝑥(𝑘) win be the same. If the initial conditions
are different, then the state 𝑥(𝑘) and the state 𝑥(𝑘) will be different. If the matrix G is a
stable one, however, 𝑥(𝑘) will approach 𝑥(𝑘) even for different initial conditions.

61
If we denote the difference between 𝑥(𝑘) and 𝑥(𝑘) as 𝑒(𝑘), or define
𝑒(𝑘) = 𝑥(𝑘) − 𝑥(𝑘)
෤
we obtain
𝑥(𝑘 + 1) − 𝑥(𝑘
෤ + 1) = 𝐺[𝑥(𝑘) − 𝑥(𝑘)]
෤
or
𝑒(𝑘 + 1) = 𝐺𝑒(𝑘).
If matrix G is a stable matrix, then 𝑒(𝑘) will approach zero and 𝑥(𝑘) ෤ will approach
𝑥(𝑘). However, the behavior of the error vector, which depends solely on matrix G, may
not be acceptable. Also, if matrix G is not a stable matrix, then the error 𝑒(𝑘) will not
approach zero. It is therefore desirable to modify the dynamic model.
It is noted that although the state 𝑥(𝑘) may not be measurable the output 𝑦(𝑘) is
measurable. The dynamic model does not make use of the measured output 𝑦(𝑘).
The performance of the dynamic model can be improved if the difference between the
measured output 𝑦(𝑘) and the estimated output 𝐶𝑥(𝑘) is used to monitor the state
𝑥(𝑘), that is, if the dynamic model is modified into the following form:
𝑥(𝑘
෤ + 1) = 𝐺 𝑥(𝑘)
෤ + 𝐻𝑢(𝑘) + 𝐾𝑒 [𝑦(𝑘) − 𝐶 𝑥෤ (𝑘)]
where matrix 𝐾𝑒 serves as a weighting matrix.

62
 𝑥(𝑘
෤ + 1) = 𝐺 𝑥(𝑘)
෤ + 𝐻𝑢(𝑘) + 𝐾𝑒 [𝑦(𝑘) − 𝐶 𝑥෤ (𝑘)]

In the presence of discrepancies between the G and H matrices used in the model and those
of the actual system, the addition of the difference between the measured output and the
estimated output will help reduce the differences between the dynamic model and the actual
system.
State observation for state-feedback control;

State feedback control system

63
 The system equations are
𝒙((𝑘 + 1)𝑇) = 𝑮𝒙(𝑘𝑇) + 𝑯𝒖(𝑘𝑇)
𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)
𝒖(𝑘) = −𝑲𝒙(𝑘)
where
x(kT) = state vector (n-vector) at kth sampling instant
u(kT) = control signal (r-vector) at kth sampling instant
y(kT) = output vector (m-vector) at kth sampling instant
G = n x n matrix
H = n x r matrix
C = m x n matrix
K = state feedback gain matrix (n x r matrix)
We assume that the system is completely state controllable and completely observable, but
𝑥(𝑘) is not available for direct measurement. The observed state 𝑥(𝑘) is used to form the
control vector u(k), or 𝑢(𝑘) = −𝐾 𝑥෤ (𝑘).
𝑥(𝑘
෤ + 1) = 𝐺 𝑥(𝑘)
෤ + 𝐻𝑢(𝑘) + 𝐾𝑒 [𝑦(𝑘) − 𝐶 𝑥෤ (𝑘)]

where 𝐾𝑒 is the observer feedback gain matrix (an 𝑛 𝑥 𝑚 matrix).

64
This last equation can be modified as
෥(𝑘 + 1) = 𝑮 − 𝑲𝒆 𝑪 𝒙
𝒙 ෥(𝑘) + 𝑯𝒖(𝑘) + 𝑲𝒆 𝒚(𝑘)

The state observer given by this equation is called a prediction observer, since the estimate
𝒙(𝑘 + 1) is one sampling period ahead of the measurement 𝒚(𝑘). The eigenvalues of
(G – 𝑲𝒆 C) are commonly called the observer poles.

Error Dynamics of the Full-Order State Observer.

Notice that if 𝑥෤ (𝑘) = 𝑥(𝑘),then
෥ 𝑘 + 1 = 𝑮෥
𝒙 𝒙(𝑘) + 𝑯𝒖(𝑘)
which is identical to the state equation of the system. Thus, if 𝑥෤ (𝑘) = 𝑥(𝑘), then the
response of the state observer system is identical to the response of the original system.
Let obtain the error equations as;
𝑥(𝑘 + 1) − 𝑥(𝑘 ෤ + 1) = 𝑮 − 𝑲𝒆 𝑪 [𝑥(𝑘) − 𝑥(𝑘)]
෤
Now let us define the difference between 𝑥(𝑘) and 𝑥(𝑘)]
෤ as the error 𝑒(𝑘):
𝑒(𝑘) = 𝑥(𝑘) − 𝑥(𝑘)
෤
Then the error dynamics becomes,
𝑒(𝑘 + 1) = 𝑮 − 𝑲𝒆 𝑪 𝑒(𝑘).

65
➢ It is seen that the dynamic behavior of the error signal is determined by the
eigenvalues of 𝑮 − 𝑲𝒆 𝑪 .

➢ If matrix 𝑮 − 𝑲𝒆 𝑪 is a stable matrix, the error vector will converge to zero for any
initial error 𝒆(0).That is, 𝑥෤ (𝑘) will converge to 𝑥(𝑘) regardless of the values of
𝑥෤ (0) and 𝑥 𝑘 .

➢ If the eigenvalues of 𝑮 − 𝑲𝒆 𝑪 are located in such a way that the dynamic behavior
of the error vector is adequately fast, then any error will tend to zero with adequate
speed.

➢ One way to obtain fast response is to use deadbeat response. This can be achieved if
all eigenvalues of 𝑮 − 𝑲𝒆 𝑪 are chosen to be zero.

66
67
Example: Consider the system
𝒙((𝑘 + 1)𝑇) = 𝑮𝒙(𝑘𝑇) + 𝑯𝒖(𝑘𝑇)
𝒚(𝑘𝑇) = 𝑪𝒙(𝑘𝑇)
where

Let design a full-order state observer. The desired eigenvalues of the observer matrix are
𝑧 = 0.5 + 𝑗0.5, 𝑧 = 05 − 𝑗0.5
and so the desired characteristic equation is
𝑧 − 0.5 − 𝑗05 𝑧 − 05 + 𝑗0.5 = 𝑧 2 − 𝑧 + 0.5 = 0
Before we proceed further, let us examine the observability matrix,
0 1
𝑪∗ 𝑮∗ 𝑪∗ = Its rank is 2.
1 −1
Hence, the system is completely observable and determination of the desired observer
feedback gain matrix is possible.

68
Referring
𝑒(𝑘 + 1) = 𝑮 − 𝑲𝒆 𝑪 𝑒(𝑘).
the characteristic equation of the observer becomes
𝒛𝑰 − 𝑮 + 𝑲𝒆 𝑪 = 0
Let us denote the observer feedback gain matrix 𝑲𝒆 , as follows:
𝑘
𝑲𝒆 = 1
𝑘2
Then the characteristic equation becomes

which reduces to
𝑧 2 + 1 + 𝑘2 𝑧 + 𝑘1 + 0.16 = 0
Since the desired characteristic equation is
𝑧 2 − 𝑧 + 0.5 = 0
by comparing these two last equations, we obtain
0.34
𝑘1 =0.34, 𝑘2 = −2 𝑲𝒆 =
−2

69