1/11/2016

Impulse response and weighting function

Chapter 2: Linear systems and
their models

• Impulse response and weighting function

• Transfer function matrices
• Transfer operator
• Input-output-models
• State-space representation g the weighting function, which for linear systems is equal
to the (unit) impulse response.
• Frequency functions
In multivariable case g is p x m-matrix, where the
• (Discrete-time systems) element (k, j) is the input response in channel k, when
the impulse enters in input channel j.

1 2

Impulse response and weighting function, cont.. Example:

Transfer function matrix

2 inputs, 1 output

System at rest at time t=0 (initial conditions zero)

At time 0 in channel 1 a unit step and in channel 2
unit impulse. The output becomes
SISO: dim(B)=n, dim(A)=m

m > n: G is ”strictly proper”

m = n: G is ”proper” strictly proper Notice also the
m < n: G is ”non-proper” final value theorem
3 4

1
 1/11/2016

Input-output representations
Transfer operator (derivative operator)

compare to discrete-time systems

the corresponding differential equation is

Note the ”connection” between the operator p in time
domain with the laplace-variable s.
Same in discrete time (operator q and variable z).

5 6

State-space representation Example: a heat exchanger

If D = 0, the system is ”strictly proper”, otherwise

”proper”

7 8

2
 1/11/2016

From one representation to another Ex. Heat exchanger

1. From state space to transfer function:

Substituting p with s gives the transfer function matrix

9 10

Canonical forms The MIMO case is much more difficult and general
algorithms complicated.

SISO-model Esim. MISO (many inputs, one output)

Realizations:

Realizations from inputs1 and 2

to output: same A- and C-mat-
controllable canonical observable canonical rices!
form form
11 12

3
 1/11/2016

Of course Matlab can be used, Control System Properties of linear systems

Toolbox commands

• Solution of the state equation

ss, tf
ss2tf • Controllability and observability
tf2ss • Poles and zeros
impulse
step • Stability
• Frequency response and frequency functions
For example. Gtf=tf(1,[1 2 3]);
Gss=ss(Gtf);

13 14

Solution of the state equation Change of the state variable

T is an invertible square matrix

new realization

Note. A corresponding discrete-time system can be

derived from this solution by assuming that the control The similarity transformation; matrices A and TAT -1
signal remains constant between sampling instants (ZOH=
zero order hold). are similar.

15 16

4
 1/11/2016

For a diagonal matrix The i:th component of the solution is

in which Bi is the i:th row of matrix B. The components

it holds xi do not depend on other components of x.
xi corresponds to the mode of

If Ci is the i:th column of the matrix C , it follows

The output can be seen as a weighted sum of the modes.

17 18

Controllability and observability

State is controllable, if there exist a control that
drives the system in a finite time from the
• Structural properties: stability, controllability state x* to the origin of the state space (note
(reachability), observability the mistake in the textbook)
• Describe how the system output depends on inputs
The system is controllable, if all states are controllable.
and internal ”states”
• Kalman, 1960’s (pioneering period of modern control State is non-observable, if
engineering)

The system is observable , if it lacks non-observable states.

19 20

5
 1/11/2016

The non-observable states form the kernel of the linear map

The controllable states are the range of the linear map
given by the controllability matrix

(observability matrix)
The system is controllable, if the rank of the
controllability matrix is full.

In the SISO case S(A,B) is a square matrix, which must The system is observable, when the observability matrix
have a non-zero determinant, in order the system to be has full rank. For a square matrix the determinant must
controllable. More generally the rank of the matrix be non-zero.
must be checked.

21 22

Note: Pole-placement:
•If the realization can be transformed to the controllable
canonical form, the system is controllable.
•If the realization can be transformed to the observable
canonical form, the system is observable.
•A SISO-system is both controllable and observable, when
there are no pole-zero cancellations in the calculation of
the transfer operator (transfer function).
mxn-matrix L can be found such that arbitrary eigenvalues
•When a state-space representation is both controllable and are obtained, if and only if
observable, it is the minimal realization of the system;
there are no realizations of lower degree that would has full rank
generate the same input-output behaviour.

23 24

6
 1/11/2016

Stabilizability and Detectability

Correspondingly, the nxp-matrix K can be found such that
the matrix A-KC has arbitrary eigenvalues, if and only if If the system is not controllable, still the controllable
modes (eigenvalues) can be influenced by state feedback;
non-controllable modes cannot be influenced.

But if non-controllable modes are (asymptotically) stable

the system is called stabilizable.

Definition: The system (A,B,C) is stabilizable, if there

has full rank. exists a matrix L such that A-BL is stable (all
eigenvalues in the stability region = left half plane or inside
the unit circle). The system is detectable, if there exists
a matrix K such that A-KC is (asymptotically) stable.

25 26

Poles and zeros To form realizations for MIMO systems is generally

difficult. It would be nice to have a method to calculate
The eigenvalues of the system matrix A are important in the the poles directly from the transfer function matrix G(s).
charcterization of the system behaviour (modes).
Note. SISO-systems do not have such problems; when
linear combinations all possible cancellations have been made, the pole
polynomial is the denominator of the transfer function.
Definition: The poles are the eigenvalues of A, where A is
the system matrix of the minimal realization. The Matlab: Gss = minreal(ss(Gtf));
dimension of a pole corresponds to the multiplicity of the
(starting from the transfer function matrix Gtf forms a
corresponding eigenvalue. The pole polynomial is the
state-space realization and then the minimal realization)
characteristic polynomial of A.
But programs are only programs!

27 28

7
 1/11/2016

Theorem: The pole polynomial is the least common

denominator of all (not identically zero) minors. The poles
The minors are obtained as determinants of the sub- of the system are the zeros of the pole polynomial.
matrices.
Ex. The matrix
has nine minors Ex.

Minors:
1, 2, 3, 4, 5, 6

The largest minors correspond to the largest sub- Pole polynomial

determinant (in the example the 2x2-cases) Poles: -2 and –1 (multiplicity two)

29 30

The realization can now be formed

g11=tf(2,[1 1]);
g12=tf(3,[1 2]);
The order of the system is three (number of state variables) g21=tf(1,[1 1]);
which leads to g22=tf(1,[1 1]);
Gtf=[g11 g12;g21 g22]

Gss1=ss(Gtf)

Gss2=minreal(ss(Gtf))

(the same result in this example case)

31 32

8
 1/11/2016

Note that the poles are the denominators of the transfer Definition:
functions in the transfer function matrix. ”Minor analysis” The system zeros (transmission zeros) are those s, for
is needed in the determination of pole multiplicities, which which the rank of the matrix
are again needed to form the minimal realization.
Zeros
The zero of a SISO-system is such s, which makes the
value of the transfer function zero (to lose rank in the
multivariable case). The zeros of a square matrix G(s) are drops (is not full). The polynomial with these zeros
the poles of s, is the zero polynomial.

For the system with equal number of inputs and outputs,

the zero polynomial is det M(s). In other cases the zeros
can be determined directly from the transfer function
matrix according to the following theorem.

33 34

Theorem: Form the maximal minors of G(s) normalized

such that the denominators contain the pole polynomial.
The zero polynomial
The zero polynomial of the system is the greatest common
divisor of these. The zeros are the zeros of this and the system has one zero s = 1.
polynomial.
Note.

Ex.

The maximal ”minors”

The pole polynomial is (verify!) –s+1
The pole polynomial (from the previous example)
which is already in the denominator

35 36

9
 1/11/2016

Stability of linear systems Result: A linear time-invariant system is BIBO-

stable, if and only if its poles are in the stability region.
Definition: The stability region of a continuous system is
the LHP = left half plane, imaginary axis excluded. For a Result: A linear time-invariant system is asymptotically
discrete-time system the corresponding region is the inside stable, if and only if the eigenvalues of the system matrix
of the unit circle. A are in yhe stability region. If the system is stable,all
eigenvalues are in the stability region or on its boundary.
For linear systems
Definition: The system is non-minimum phase, if
and so it has at least one zero outside the stability region.
Otherwise the system is minimum phase.
BIBO-stable, if

37 38

Frequency response and frequency How to define the gain of a MIMO-system?

functions
Consider the transfer function matrix

The frequency response (Bode) of each channel separately

does not give full information about the behaviour of the
system (interconnections; different ”directions” of the
the input in channel k (other input channels zero) multi-channel input function .

What about the eigenvalues of G? But they are defined

output in channel j only for square systems, and do not generally give a
reliable view about the gain of a multivariable system.

39 40

10
 1/11/2016

Solution: Singular values

A=[1+i;2-3i]
A=
Consider the mapping
1.0000 + 1.0000i
2.0000 - 3.0000i

A'
A p x m- matrix 1.0000 - 1.0000i 2.0000 + 3.0000i

x m x 1-vector A.'
1.0000 + 1.0000i 2.0000 - 3.0000i
y p x 1-vektor
Matlab: ’ means actually taking the konjugate transpose.
(complex values are allowed) For an ordinary transpose,write .’
Definition: The hermitian matrix (adjoint) of A, or For real matrices the shorter form is of course used also
is obtained by taking the transpose and then the complex in the case of an ordinary transpose.
conjugate of each term
Ex.

41 42

The familiar rules are valid, for example. Correspondingly, A is negative definite, if
x*Ax < 0 (often written as A < 0)
Definition: A complex-valued matrix is hermitian,
(self-adjoint), if A is positive semidefinite, if

(cf. symmetric for real matrices)

A is negative semidefinite, if
Hermitian (and symmetric) matrices have real eigenvalues.

Definition: Let A be a hermitian matrix. It is Note that if A is hermitian, the square form x*Ax
positive definite, if the scalar x*Ax > 0, for all is always real.
non-zero vectors x.
Result: A hermitian A is [Link]., when all eigenvalues are
x*Ax > 0 (often written as A > 0) positive; pos. semidefinite, when all eigenvalues are
non-negative.
43 44

11
 1/11/2016

Corresponding results are valid also in the case of negative Let us return to study the gain of the map
definite matrices.
How ”big” is y when compared to x?
To check the positive definitness of a symmetric, real
matrix: the Sylvester rule. Well, the 2-norm of x is

So
The matrix is hermitian, eigenvalues real

; largest and smallest

45 46

Then (Rayleigh-Ritz The induced matrix norm corres-

inequality) ponding to 2-vector norm
and the definition follows
The singular values are generated naturally by the
Definition: The singular values of A are singular value decomposition (SVD):
in which the values are the eigenvalues of
Result: For any real or complex matrix A there always
the largest eigenvalue is denoted as
exists a factorization
and the smallest one as
(SVD)
When y = Ax, then
A (n x m), U (n x n), V (m x m)
The gain of the matrix is between the smallest and U and V unitary
largest eigenvalue. It is also a norm.
(cf. orthogonal in the case of real matrices)

47 48

12
 1/11/2016

(n x m) is a real matrix, and the singular values of A

are located in the main diagonal in descending order.
If A is complex, U and V are also complex; otherwise The gain of the system at freq. is between the
real. smallest and largest singular value of

The columns of U and V are the unit eigenvectors of Corresponding to Bode diagrams in SISO systems
AA* and A*A, respectively. They represent the output the singular values are plotted as functions of frequency;
and input directions. the gain in all ”directions” is between the smallest and
largest singular value.
Matlab: Norm of the frequency function:
Frequency functions: (the largest singular value)

is the system gain

49 50

Ex. The calculation of singular values is difficult.

In Matlab: the sigma command is helpful.

A=[-0.21 0.2;0.2 -0.21];

B=0.01*eye(2);
C=eye(2);
D=zeros(2,2);
sigma(A,B,C,D)

51 52

13
 1/11/2016

What about ”directions” ? Result:

Transfer function 0.7071 0.7071 input directions

=V
-0.7071 0.7071

0.0006 0 singular values

=D
Look at the frequency s = 0 0 1.0000 0.0006 and 1.0
Singular value is related to the direction
g0=[0.21 0.2;0.2 0.21]/0.41;
and to
[V,D]=eig(g0'*g0)
The input form has a very minor influence
to the output

53 54

Useful Matlab-commands:

ss2zp, zp2ss, tf2zp, zp

tzero
pole
pzmap
lsim
eig, roots
bode
sigma
obsv, ctrb

55

14